CN1191717C - Method and device for reducing blocking artifact in the DCT domain - Google Patents

Abstract

A method and device for removing blocking artifacts in the frequency domain. Two frequency domain blocks (A & B) are received in a video stream. A third frequency domain block is computed directly in the transform domain which overlaps the boundary between the first and second frequency domain blocks. The coefficients of the third block are modified in the frequency domain to remove blocking artifacts and then this modification is translated into a modification of the two frequency domain blocks (A and B) in the frequency domain.

Description

In discrete cosine transform domain, reduce the method and apparatus of block noise

The present invention relates to a kind of block noise (blocking artifacts) that reduces in the view data, be meant especially, reduce the method and apparatus of the block noise of the view data that only is used for frequency domain.

The compression of image makes image and to use than compressed format data flow still less not with a kind of form of coding to transmit in communication channel.Discrete cosine transform (DCT) coding is a kind of compression scheme that is widely known by the people.Image is divided into a plurality of little rectangular region or " piece ".Each piece is transformed coding and transmits by communication channel.At receiving terminal, these pieces are decoded and be reduced into original image.Typical block structure is the form of 8X8 pel array.Each piece all will quantize then through dct transform.

DCT is a kind of linear transformation mode, and produces a new block of pixels; Each new pixel all is a linear combination of the input pixel of original picture block.Image block coding techniques based on DCT can make the reception image quality decrease, shows as block noise in form.When piece image during by block encoding, the piece of reconstruction can see on the image of rebuilding but the spectator can see the border of piece, and this is owing to uncorrelated quantizing noise causes usually.Quantizing noise is owing to producing step on saltus step of generation or the border at piece between the piece.Usually compression ratio is big more, and block noise is just big more.Blocking effect is very obvious for the spectator, even allows the people be difficult to stand, because human eye is very responsive to " step " of block boundary.

This problem has been found and existing multiple effort is attempted to address this problem.Block noise is owing to discontinuity precipitous between adjacent block produces, and eliminates these discontinuities and just can reduce block noise.Many methods in the past all are to handle in the scope of space (pixel).Wherein a kind of technology is in pixel domain, uses a space variable low pass filter on the border of piece.The problem of this method is the decompression that also comprised image before using filter.Though for the compressed image of certain form of storage of great use, need compress once more image by this method after reducing block noise for this method.

Provide a kind of method and apparatus that reduces block noise at this, by analyzing among the DCT characteristic on border between two pieces, and at frequency domain discontinuity precipitous on the block boundary is become smoothly to reduce block noise.First area block (A) and second area block (B) are received by video flowing.Calculate the 3rd area block (C) of first and second area blocks that cover the border between first and second area blocks then.The 3rd piece is to make the discontinuity of A piece and B block boundary become level and smooth by the DCT coefficient of adjusting the 3rd piece.Variable after the coefficient calculations of the 3rd piece will be converted into first and second pieces (A, the variable of coefficient B).

The objective of the invention is only to realize this smoothing processing at frequency domain by the 3rd piece C.

The variable that the objective of the invention is first and second pieces should satisfy equation δ C=δ AM ₁+ δ BM ₂

Here δ C is the variable of the 3rd piece, and δ A and δ B are each independents variable of first and second pieces.M ₁And M ₂It is the known matrix of value separately.

Another object of the present invention is by selecting δ A and δ B that the variable of first and second pieces is minimized.

Further purpose of the present invention is to make the value of δ A opposite with the value of δ B.

Another object of the present invention is whether decision has consistent relatively pixel value in first and second pieces when the pixel value of two piece adjacent boundaries produces transition.

Another purpose of the present invention is when piece A and piece B have the pixel value of relative unanimity they not to be handled.This purpose is to realize by the big DCT coefficient that reduces the 3rd piece C medium-high frequency composition.

Another purpose of the present invention is the coefficient that changes the 3rd piece C by the big coefficient of the high frequency composition that reduces the 3rd piece C, if they are almost nil in the corresponding coefficient of first and second pieces.

Other purpose of the present invention and advantage will be partly more directly perceived and apparent in specification.The present invention comprises several steps, the correlation between these steps, and with the construction features of these step corresponding device thereof, the configurations of each assembly etc. below will be described in detail for example to all these contents.

Protection scope of the present invention will be illustrated in claims.

Embodiments of the invention all are by analyzing borderline DCT characteristic between two pieces, and at frequency domain discontinuity precipitous on the block boundary are become smoothly to reduce block noise.In video flowing, receive two area block A and B, calculate the 3rd the piece C that covers the border between two area block A and the B then.The effect of the 3rd piece is to make that borderline discontinuity becomes smoothly between piece A and the piece B, and its method is the correction of revising the DCT coefficient of the 3rd piece and this correction being converted to the coefficient of piece A and B.The reducing of block noise illustrates with reference to embodiment, and the piece A and the B of reference in first embodiment consistent pixel value is arranged in each comfortable spatial domain, but the borderline pixel value of interblock have a step or saltus step; The piece A and the B of reference in second embodiment might not have consistent pixel value in spatial domain.

Fig. 1 a represents two 8 * 8 block of pixels ' a ' and ' b ', and the 3rd block of pixels ' c ' that is superimposed upon piece ' a ' and ' b ' has covered border ' e '.Fig. 1 b is the element a that illustrates in the piece in the first embodiment of the invention ' a ' _IjAnd the element b in the piece ' b ' _IjIn this embodiment, a _Ij(i, value j) is the same, in this case, for i=07 and j=07, is expressed as θ, i.e. a at all ranks _Ij=θ (similarly, works as b _IjAll ranks (i, value j) is the same, in this case,, be represented as for i=07 and j=07 ↓, i.e. b _Ij=↓), after this, a _IjAnd b _IjRepresent an independent pixel value separately.When compression ratio is very high, step will appear in the border ' e ' between piece ' a ' and ' b ', for example left side of ' e ' and right side on the border, all pixel value a _IjAnd b _IjAll be constant, but on border ' e ', a very big discontinuity arranged.Fig. 2 a represents how the pixel value in the piece ' c ' changes.By seeing the edge that a step is arranged among Fig. 2 a on 4.This step edge will cause block noise.In order to eliminate this block noise, step edge must be become the edge on a slope to make it level and smooth, shown in Fig. 2 b.Purpose of the present invention makes the DCT of piece ' c ' exactly, and promptly C becomes smoothly at frequency domain, just can eliminate the middle step of piece ' c ', more this level and smooth variation is converted to the variation of the DCT coefficient of piece A and B.

When receiving a bit stream that compressed with JPEG or MPEG, through just can from the bit stream that compresses, obtaining DCT piece A and B behind the variable-length decoding, but can not obtain DCT piece C.Therefore, in order to make the DCT piece C level and smooth DCT coefficient that just must computing block C that becomes.This calculating not only comprises the DCT value of merging piece A right half and the DCT value of piece B left half, because each DCT piece all is linear combinations of all input pixels in the piece, also will carry out very complicated calculating to DCT piece C.A kind of method of seeking the DCT coefficient of piece C is that piece A and B are transformed in the spatial domain, and comes computing block C by all pixels of piece ' c ' are carried out linear combination.This is a kind of very loaded down with trivial details method and need decompresses to piece.Do not need to carry out dct transform among the present invention and just can obtain DCT piece C by DCT piece A and B.

Because piece ' a ', ' b ' and ' c ' can regard as simply by pixel value a _IjAnd b _IjConstitute the matrix of element, therefore, matrix ' c ' can be write as the following form with matrix ' a ' and ' b ' expression:

c＝aK ₁+bK ₂ (1)

Here K ₁And K ₂It all is 8 * 8 matrix.Matrix K ₁Knownly be:

${\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="C00803167.300171.png"\; state="\{\{state\}\}">K</figure-callout>}}_1=\left[\begin{array}{cc}0& 0\\ I& 0\end{array}\right]$

Here 0 is that an all values all is 04 * 4 matrixes, and I also is one 4 * 4 matrix, and all values on its diagonal all equals 1, and off-diagonal all values all equals 0.K ₂=K ₁ ^T(be K ₂Be K ₁Transposed matrix).The DCT of piece ' c ', promptly C is to be obtained by following DCT matrix multiple.Be known as the DCT that obtains a matrix, must have multiplied each other by following mode:

A＝DaD ^T

B＝DbD ^T

C＝DcD ^T (2)

(D is the DCT matrix among Fig. 3 here, D ^TBe the transposed matrix of matrix D shown in Figure 8).

Original matrix (' a ', ' b ', ' c ') can be gone out by the DCT matrix computations in order to following equation:

a＝D ^TAD (3)

b＝D ^TBD (4)

c＝D ^TCD

' c ' in the equation 2 replaced with equation 1:

C＝D(aK ₁+bK ₂)D ^T (5)

And, obtain ' a ' in the equation 5 and ' b ' usefulness equation 3 and 4 replacements:

C＝D(D ^TADK ₁+D ^TBDK ₂)D ^T

＝(DD ^TADK ₁+DD ^TBDK ₂)D ^T

＝DD ^TADK ₁D ^T+DD ^TBDK ₂D ^T (6)

Known DD ^T=D ^TD=I

C＝ADK ₁D ^T+BDK ₂D ^T (7)

This equation can also be write as:

C＝AM ₁+BM ₂ (8)

By equation 7 as can be seen, M ₁And M ₂Be fixed value and do not rely on matrix ' a ', ' b ' and ' c '.

Each representing matrix M of Fig. 4 a and Fig. 4 b ₁And M ₂By these matrixes M as can be seen ₁With M ₂The value of the 1st, 3,5,7 elements is identical in the odd-numbered line, and the value of the 2nd, 4,6,8 element is opposite.Then opposite in even number line.Therefore, M ₁With M ₂Can be write as following form:

M ₁＝Com+Dif (9)

M ₂＝Com-Dif (10)

Here Com only comprises M ₁With M ₂In identical element, remaining element all is 0, Dif then only comprises different element (all the other elements all are 0), each is shown in Fig. 5 a and 5b.With equation 9 and 10 M that replace in the equation 8 ₁And M ₂, the DCT Matrix C can obtain (not needing conversion to return spatial domain) by following mode from DCT matrix A and B:

C＝A(Com+Dif)+B(Com-Dif)

＝(A+B)Com+(A-B)Dif (11)

Because Matrix C om only comprises identical element, it has and is less than 32 nonzero element.Same, matrix D if also has only the element of 32 non-zeros.Calculate C without equation 8 and can simplify calculating with equation 11, although because equation 8 and 11 all needs with two matrix multiples, it is zero that more element is arranged in the matrix that equation 11 multiplies each other.

Therefore do not need piece A and B are transformed to pixel domain, just can obtain the DCT coefficient of the dct transform of piece ' c ' by equation 11.

In case the DCT coefficient calculations of piece C is finished, this piece just must be positioned at the middle step edge of piece to eliminate by " smoothing processing " in the spatial domain.For piece C being become smoothly, and do not need conversion again to return pixel domain, when the piece ' c ' of spatial domain becomes the slope shape edge by step shape edge, will make analysis the variation of the coefficient of piece C in the DCT territory.

Fig. 6 a represents a picture element matrix ' c ', and it produces a step at the 4th row and the 5th row.Fig. 6 b represents a matrix newc, has carried out an interpolation in the pixel value of its matrix ' c ' in Fig. 6 a.Newc is the piece after level and smooth in pixel domain, and the edge in its piece is a slope shape.The 1st row comprise the pixel value a of piece ' a ' in Fig. 6 b _Ij, last row comprise the pixel value b of piece ' b ' _IjThe pixel value of 2-7 row changes from left to right by the mode of linear interpolation.Can find out the edge that between the 4th row and the 5th row, no longer includes a step by Fig. 6 b.Obviously linear interpolation is not the method for this step edge of unique removal, is to adopt this method for convenience of explanation here.

The DCT form of piece newc is NEWC, shown in Fig. 7 c.The DCT form of piece ' c ' is C, shown in Fig. 7 a.The different coefficients of first row that only are of piece C and ' c ' are corrected.Because the dct transform of a piece can pass through DcD ^TFormula obtains, and matrix D and D have been shown among Fig. 3 and 8 ^T, matrix ' c ' has been shown among Fig. 6 a, first row of piece C can be write out by following equation.(notice that all remaining elements all are zero among the piece C, here a _Ij=θ is to i=07, all values of j=07, b _Ij=↓ to i=07, the j=07 all values)

C ₀₀＝4(θ+↓)，C ₀₁＝2 ^1/22.563(θ-↓)，C ₀₃＝2 ^1/20.9(θ-↓)，

C ₀₅＝2 ^1/20.601(θ-↓)，C ₀₇＝2 ^1/20.511(↓-θ)

(above-mentionedly be used to calculate C ₀₀-C ₀₇Equation can be used for separately a _IjAnd b _IjAll the same any two pieces of pixel value.)

With ' c ' in the replacement of the newc among Fig. 4 b equation 2

DnewcD ^T＝NEWC

First row of NEWC can be write (all surplus elements all are zero among the same NEWC):

NEWC ₀₀＝4(θ+↓)，NEWC ₀₁＝2 ^1/21.841(θ-↓)，NEWC ₀₃＝2 ^1/2.1918(θ-↓)，

NEWC ₀₅＝2 ^1/20.0576(θ-↓)，NEWC ₀₇＝2 ^1/20.50139(θ-↓) (12)

Formula 12 need be used the value θ of spatial domain _IjAnd _Ij, but these values are easy to be come out by the coefficient calculations of piece A and B, and do not need to be transformed in the spatial domain.In the present embodiment, to all i=07 and j=07, a _Ij=θ=40 and b _Ij=↓=70; But θ and ↓ value depend on the video format that video flowing receives.Fig. 9 A and 9B represent piece A and B (dct transform of the piece shown in Fig. 1 b ' a ' and ' b ').DCT piece A and B respectively have only a nonzero coefficient, when this coefficient is removed by 8, can obtain area block ' a ' and ' b ' pixel value θ separately with ↓.A in the present embodiment for example _Ij=40=θ is to i=07 and j=07, ↓=70=b _IjTo i=07 and j=07.Therefore only need piece A and B DCT coefficient separately just can directly to obtain the pixel value a of frequency domain divided by 8 _IjAnd b _IjFor piece C is transformed into NEWC, first row coefficient of piece C must come conversion with equation 12.So only piece C is just handled and step shape edge can be become level and smooth slope shape edge at frequency domain.

As mentioned above, piece C is not present in the video flowing, and it is the calculating of carrying out in order to eliminate block noise.Therefore must convert piece A in the video flowing and the index variation of B to the smoothing processing of piece C.By equation C=AM ₁+ BM ₂Can obtain

C+δC＝(A+δA)M ₁+(B+δB)M ₂

δC＝δAM ₁+δBM ₂

With δ C (δ C represents piece C is transformed into the difference of piece NEWC here) writing NEWC-C, then

NEWC-C＝δAM ₁+δBM ₂

Therefore any value that satisfies this equation δ A and δ B can make the border between piece A and the B become smoothly, and eliminates block noise.This equation separated many kinds, but in order to make picture quality that best effect be arranged, just should make the variable quantity minimum of each piece, and make the variation of a piece and changing in the opposite direction of another piece, so just can between piece, produce a level and smooth slope shape edge.Make δ A=-δ B in a preferred embodiment of the present invention.Then equation becomes:

NEWC-C＝δAM ₁+δAM ₂

NEWC-C＝δA(M ₁-M ₂)

(NEWC-C)(M ₁-M ₂) ^-1＝δA

δ A is shown in Figure 11.

The index variation of piece A and B (A+ δ A, B+ δ B) is respectively shown in Figure 12 and 13.By piece A and B are revised by this way, piece A and B do not need to transform to the edge that just can make in the spatial domain between them and become level and smooth.

Figure 14 shows a kind of system that uses method of the present invention and/or device.The video flowing of 10 pairs of inputs of a variable length decoder is decoded, and the DCT piece that 20 pairs of inverse quantizer has block noise calculates.Block noise is eliminated system 30 block noise is eliminated from the DCT piece, and provides the DCT piece to be used for showing or storage.

More than be when the piece in the spatial domain comprises identical pixel value, how eliminate the explanation of block noise.Employed a kind of diverse ways below will be described when the piece in the spatial domain comprises many high frequency compositions.

When not comprising identical pixel value in the piece, the first method that reduces block noise is, reduces the high frequency composition that produces owing to block boundary in the piece C.When computing block C, use the method for being explained here, will in piece C, produce a large amount of high frequency DCT compositions, this is owing to cause on the border of piece.

Be set to zero or reduce their value by high frequency coefficient value big in the piece C, just can realize smoothing processing.In case after this correction to piece C is finished, just as stated above the variation of piece C is converted to the variation of piece A and B.This method can be used for comprising the piece of same pixel, but for the piece that comprises same pixel, uses the method for linear interpolation will obtain a more accurate result.

In another embodiment of the present invention, the smoothing processing of piece C is to change indivedual coefficient values in the C to zero, when this coefficient value when corresponding value also is zero in piece A and B.The reason of Chu Liing is like this, supposes that piece A and B belong to same image-region, and then the frequency characteristic of piece A and B should be similar.Because piece C also belongs to identical image-region, then the frequency characteristic of piece C should be similar to the frequency characteristic of piece A and B also.But owing among the piece C block boundary is arranged, the high frequency characteristics that will exist some to represent block boundary among the piece C, but in piece A and B, do not exist, therefore should be set to zero.Thereby all be that zero high frequency composition is set to zero in piece A and B in piece C.This variation of piece C will be converted into the variation among piece A and the B as stated above.

Although the scope of above-mentioned discussion has been restricted to a vertical boundary between piece A and the B, be appreciated that the situation that expands to a horizontal boundary also is same being suitable for.In addition, the correction of above-mentioned DCT coefficient can iterate in image block, until obtaining required smoothed curve.

Can obviously find out by above-mentioned explanation; purpose of the present invention can effectively realize; and; owing to must have some distortion when stating method on the implementation; under the premise without departing from the spirit and scope of the present invention; all the elements that above-mentioned explanation and relevant drawings are comprised all just are used to illustrate, and can not limit protection scope of the present invention.

Claims (12)

1, a kind of method of eliminating block noise comprises:

Receive first area block of representing first block of pixels;

Receive second area block of representing second block of pixels;

Calculate the 3rd area block of representative the 3rd block of pixels that covers a border between first and second block of pixels;

Reduce the block noise on the border between first and second block of pixels by the index variation (δ C) of calculating the 3rd area block; And

Calculate the respective change (δ A, δ B) in first and second area blocks, this respective change will cause the index variation (δ C) of the 3rd area block.

2, the method for claim 1 is characterized in that the step of calculating the respective change in first and second area blocks satisfies following condition

δC＝δAM ₁+δBM ₂

Here M ₁And M ₂It is constant.

3, method as claimed in claim 2 is characterized in that δ A and δ B selects so that the variation minimum in first and second area blocks.

4, method as claimed in claim 2 is characterized in that the value of δ A is opposite with the value of δ B.

5, the method for claim 1, the step that it is characterized in that calculating the index variation (δ C) of the 3rd area block are by the big coefficient that reduces the 3rd area block medium-high frequency composition and deduct the 3rd area block and finish.

6, the method for claim 1, the step that it is characterized in that calculating the index variation of the 3rd area block is to be the coefficient of revising the 3rd area block under zero the situation in first and second area blocks by each coefficient at the 3rd area block, to reduce high frequency coefficient, deduct the 3rd area block then and finish.

7, a kind of device that reduces block noise comprises:

Receiving system is used to receive first and second area blocks of representing first block of pixels and second block of pixels respectively;

Processor is used for

A.) calculating covers the 3rd area block of representative the 3rd block of pixels on a border between first and second block of pixels,

B.) reduce the block noise on the border between first and second block of pixels by the index variation (δ C) of calculating the 3rd area block; And

C.) respective change (δ A, δ B) in calculating first and second area blocks, this respective change will cause the index variation (δ C) of the 3rd area block.

8, device as claimed in claim 7 is characterized in that, described processor calculates the respective change in first and second area blocks, makes it to satisfy following condition

δC＝δAM ₁+δBM ₂

Here M ₁And M ₂It is constant.

9, device as claimed in claim 8 is characterized in that, described processor calculating δ A and δ B are so that make the variation minimum of first and second area blocks.

10, device as claimed in claim 8 is characterized in that, the value that described processor calculates δ A is opposite with the value of δ B.

11, device as claimed in claim 7 is characterized in that, described processor by reducing the 3rd area block medium-high frequency composition big coefficient and deduct the index variation (δ C) that the 3rd area block calculates the 3rd area block.

12, device as claimed in claim 7, it is characterized in that, described processor is the coefficient of revising the 3rd area block under zero the situation in first and second area blocks by each coefficient at the 3rd area block, to reduce high frequency coefficient, deduct the index variation that the 3rd area block calculates the 3rd area block then.

Images