CN108933942A - A kind of filtering method compressing video and the filter for compression video - Google Patents
A kind of filtering method compressing video and the filter for compression video Download PDFInfo
- Publication number
- CN108933942A CN108933942A CN201710386993.4A CN201710386993A CN108933942A CN 108933942 A CN108933942 A CN 108933942A CN 201710386993 A CN201710386993 A CN 201710386993A CN 108933942 A CN108933942 A CN 108933942A
- Authority
- CN
- China
- Prior art keywords
- block
- threshold
- module
- image block
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/102—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or selection affected or controlled by the adaptive coding
- H04N19/117—Filters, e.g. for pre-processing or post-processing
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/102—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or selection affected or controlled by the adaptive coding
- H04N19/124—Quantisation
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
- H04N19/169—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the coding unit, i.e. the structural portion or semantic portion of the video signal being the object or the subject of the adaptive coding
- H04N19/17—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the coding unit, i.e. the structural portion or semantic portion of the video signal being the object or the subject of the adaptive coding the unit being an image region, e.g. an object
- H04N19/176—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the coding unit, i.e. the structural portion or semantic portion of the video signal being the object or the subject of the adaptive coding the unit being an image region, e.g. an object the region being a block, e.g. a macroblock
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/60—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/80—Details of filtering operations specially adapted for video compression, e.g. for pixel interpolation
Abstract
The present invention provides a kind of filtering methods and device for compression video.The present invention is handled by Higher-order Singular value decomposition, the quality of high compression video can be proposed, to further improve video compression performance effectively using the non local spatial correlation in video frame.Meanwhile applying filtering method of the invention in Video coding loop, the quality of reference frame can be improved, further increase forecasting efficiency, to promote coding efficiency.The present invention is while proposing the filtering method based on non local spatial correlation for compression video, it is also proposed that corresponding device.
Description
Technical field
The present invention relates to field of video image processing, and in particular to a kind of filtering method and device for compression video.
Background technique
Due to lossy compression, quantizing noise usually there will be in the decoded reconstruction image of video, including：Blocking artifact, vibration
Bell effect, blurring effect etc., these will affect video quality.The quality of reconstructed frame can be improved in loop filter, and there are two types of rings
Path filter --- de-blocking filter and the adaptive offset filter of sampling.But video image is all only utilized in both filters
Local spatial correlation, have some limitations.Adaptive filter method in coding and decoding video utilizes the non-office of video frame
Portion's correlation finds a series of similar block in the picture, and each similar block is arranged in column vector, then all similar block groups
At a similar block group, and it is denoted as the form of two-dimensional matrix.Then two-dimensional transform base-singular value decomposition is selected, to every
The two-dimensional matrix of a similar block composition is converted, and threshold process is carried out on transform domain and is rebuild by inverse transformation and weighting
Means, finally obtain filtered as a result, this method can lower quantization noise to a certain extent.Based on higher order singular
It is worth the Image denoising algorithm decomposed, first finds similar block group in the picture, the two dimensional image block in group is directly arranged in three-dimensional
Array；Then to three-dimensional array application Higher-order Singular value decomposition, image is switched back to another mistake after the processing of decomposition coefficient threshold application
Space；Finally average operation is taken to complete denoising overlaid pixel gray value.Since the above method is directed to answering for image denoising
With so the selection of threshold value is related with the variance of noise in the noisy image that estimation obtains.In compression video, quantizing noise is
Due to caused by quantization, and quantization degree is related with quantization step, and energy and the quantization step of quantizing noise have exponent relation.
Above-mentioned Higher-order Singular value decomposition is a kind of decomposition algorithm for high level matrix, it is assumed that be decomposed three of input
Tieing up tensor isThen the calculation formula of three-dimensional tensor is：
Z=S ×1(U1)×2(U2)×3(U3)
Wherein S is core tensor (coefficient matrix after decomposition),It is Z respectively
Mould n (n=1,2,3) be unfolded Z(1), Z(2), Z(3)Left singular matrix after making singular value decomposition (see Fig. 3).Find out three it is unusual
After matrix, the calculation formula of core tensor is as follows：
Wherein S(1)Matrix, Z is unfolded in the mould one for representing core tensor(1)Matrix is unfolded in the mould one for representing tensor to be decomposed,
Indicate Kronecker product, two matrixesWithKronecker product representation it is as follows：
It should be pointed out that a kind of above-mentioned mode for only calculating core tensor, also can use matrix norm to be decomposed
Two expansion or three unfolding calculation of mould, do not constitute the limitation to this patent.
Summary of the invention
The purpose of the present invention is pass through a kind of needle using the non local correlation and Higher-order Singular value decomposition in video image
To the filtering method and device of compression video, the method and device utilize the available information in encoding and decoding end, such as encoding frame type, amount
Change the adjustment that the information such as step-length carry out the filtering threshold of part, mainly for the removal of quantizing noise in compression video, Ke Yiyou
The quality of the raising video frame of effect, further promotes compression performance.
To achieve the above object, the present invention takes following technical scheme：
A kind of filtering method compressing video, including：
A.
(a) using the image block P of p × q size to be processed, K-1 similar block is found using Block- matching in reconstructed image；
Image block P is formed to three rank tensor Z, Z ∈ R together with K-1 similar blockp×q×K；Wherein Z (:,:, m) and it is image block P.
(b) according to the quantization step Qstep of image block P, the threshold tau of filtering algorithm is determined：
τ=g (Qstep)
B. Higher-order Singular value decomposition is carried out to the three rank tensor Z, obtains singular value matrix U1, U2, U3And resolving system
Matrix number S, wherein U1∈Rp×p, U2∈Rq×q, U3∈RK×K, S ∈ Rp×q×K。
C. according to the filtering threshold τ, each of S element is handled using threshold method, obtains S'：
Wherein S (i, j, k) is the resolving system numerical value in decomposition coefficient matrix S at (i, j, k), S'(i, j, k) be S (i, j,
K) after handling in S' corresponding position value.
D. with decomposition coefficient the matrix S', singular value matrix U after threshold process1, U2, U3, it is calculated according to following methods
Z'(:,:, m) and filtered as a result, wherein as image block P：
Z'=S' ×1(U1)×2(U2)×3(U3)
×nIndicate mould n tensor product, n=1,2 or 3；
A kind of filtering method compressing video, it is characterised in that：
A. for a certain location of pixels PP in image f, the middle image block { P by PP is found outt| t=1,2 ... T },
Middle T is the image block number by PP, PtIn (it,jt) correspond to PP.
B. for each image block Pt,It is handled by the following method：
(a)
● utilize p to be processedt×qtThe image block P of sizet, it is similar in reconstructed image using Block- matching to find K-1
Block；By image block PtThree rank tensor Z are formed together with K-1 similar blockt, Zt∈Rp×q×K；Wherein Zt(:,:, m) and it is image block
Pt。
● according to image block PtQuantization step Qstep, determine the threshold tau of filtering algorithm：
τ=g (Qstep)
(b) to the three rank tensor ZtHigher-order Singular value decomposition is carried out, singular value matrix (U is obtained1)t, (U2)t, (U3)t
And decomposition coefficient matrix St, wherein (U1)t∈Rp×p, (U2)t∈Rq×q, (U3)t∈RK×K, St∈Rp×q×K。
(c) according to the filtering threshold τ, using threshold method to StEach of element handled, obtain St'：
Wherein St(i, j, k) is decomposition coefficient matrix StIn resolving system numerical value at (i, j, k), St' (i, j, k) be St(i,
J, k) processing after St' in corresponding position value.
(d) the decomposition coefficient matrix S after threshold process is utilizedt', singular value matrix (U1)t, (U2)t, (U3)tAccording to following
Z is calculated in methodt'(:,:, m) and it is used as image block PtIt is filtered as a result, wherein：
Zt'=St'×1(U1)t×2(U2)t×3(U3)t
×nIndicate mould n tensor product, n=1,2 or 3；
C. to { Pt' | t=1,2 ... T in (it,jt) pixel value of position merged, as position PP in the image
Filter result.
Further, the g (Qstep) is following one：
● g (Qstep)=aQstep+b
● g (Qstep)=c (Qstep)d
Wherein coefficient value a, b, c, d is obtained according to encoding frame type, bright chroma component type, 0 < a≤0.2,0 b≤1 <,
0 < c≤1,0 d≤1 <.
Further, coefficient value a, b, c, d of chromatic component U/V are identical, different from the coefficient value of luminance component Y.
The threshold tau of luminance component is greater than the threshold tau of chromatic component；The threshold tau that the threshold tau of I frame is greater than B frame is greater than the threshold tau of P frame.
A kind of filter for compression video, it is characterised in that including：
A.
● similar block finds module, and input includes the image block P and video reconstruction image f of p × q size to be processed,
It includes three rank tensor Z, Z ∈ R of similar block that it, which is exported,p×q×K.The function of the module is, with image block P, Block- matching to be utilized in f
Method finds K-1 similar block；Image block P is formed to three rank tensor Z together with K-1 similar block；Wherein Z (:,:, m) and it is figure
As block P.
● threshold calculation module, input include quantization step Qstep, and output includes filtering threshold τ.The function of the module
It can be, according to the quantization step Qstep, to calculate the threshold tau of filtering algorithm.
B.HOSVD conversion module, input include three rank tensor Z of similar block, and output includes singular matrix U1, U2, U3With
And decomposition coefficient matrix S, wherein U1∈Rp×p, U2∈Rq×q, U3∈RK×K, S ∈ Rp×q×K.The function of the module is, to described
Three rank tensor Z of similar block carries out Higher-order Singular value decomposition.
C. threshold process module, input include the threshold tau and the decomposition coefficient matrix S, output packet
Decomposition coefficient matrix S' after including threshold process.The function of the module is to carry out to each of decomposition coefficient matrix S element
Following threshold process, obtains S'：
Wherein S (i, j, k) is the resolving system numerical value in decomposition coefficient matrix S at (i, j, k), S'(i, j, k) be S (i, j,
K) corresponding position value in S' after handling.
C.HOSVD inverse transform block, input include singular matrix U1, U2, U3, and treated decomposition coefficient matrix
S', output include the filtered result Z'(of image block P:,:,m).The function of the module is, with S', U1, U2, U3, according to
Z'(is calculated in lower method:,:, m) and filtered as a result, wherein as image block P：
Z'=S' ×1(U1)×2(U2)×3(U3)
Wherein, ×nIndicate mould n tensor product, n=1,2 or 3；
The filter of the compression video, it is characterised in that the similar block finds tri- rank of output-of module
Amount Z is connected with the input of HOSVD conversion module.The output of the HOSVD conversion module-decomposition coefficient matrix S and threshold value
The input of processing module is connected, and output-threshold tau of threshold calculation module is connected with threshold process module.The HOSVD becomes
Change the mold output-singular matrix U of block1, U2, U3It is connected with the input of HOSVD inverse transform block, the output-of threshold process module
Treated, and decomposition coefficient matrix S' is connected with the input of HOSVD inverse transform block.
A kind of filter compressing video, it is characterised in that including：
A. image block determining module, input includes a certain location of pixels PP in video reconstruction image f and f, defeated
It out include by the image block { P by PPt| t=1,2 ... T }, wherein T is the image block number by PP, PtIn (it,
jt) correspond to PP.The function of the module is to find out all image blocks for passing through location of pixels PP in image f.
B. for each image block Pt,It is handled using following apparatus：
(a)
● similar block finds module, and input includes p to be processedt×qtThe image block P of sizetWith video reconstruction image
F, output include three rank tensor Z of similar blockt, Zt∈Rp×q×K.The function of the module is, with image block Pt, block is utilized in f
Matched method finds K-1 similar block；By image block PtThree rank tensor Z are formed together with K-1 similar blockt；Wherein Zt
(:,:, m) and it is image block Pt。
● threshold calculation module, input include quantization step Qstep, and output includes filtering threshold τ.The function of the module
It can be, according to the quantization step Qstep, to calculate the threshold tau of filtering algorithm.
(b) HOSVD conversion module, input include three rank tensor Z of similar blockt, output includes singular matrix (U1)t,
(U2)t, (U3)tAnd decomposition coefficient matrix St, wherein (U1)t∈Rp×p, (U2)t∈Rq×q, (U3)t∈RK×K, St∈Rp×q×K.It should
The function of module is, to the three rank tensor Z of similar blocktCarry out Higher-order Singular value decomposition.
(c) threshold process module, input include the threshold tau and the decomposition coefficient matrix St, output
Including the decomposition coefficient matrix S after threshold processt'.The function of the module is, to decomposition coefficient matrix StEach of element
Following threshold process is carried out, S is obtainedt'：
Wherein St(i, j, k) is decomposition coefficient matrix StIn resolving system numerical value at (i, j, k), St' (i, j, k) be S (i,
J, k) processing after St' in corresponding position value.
(d) HOSVD inverse transform block, input include singular matrix (U1)t, (U2)t, (U3)t, and treated decompose
Coefficient matrix St', output includes image block PtFiltered result Zt'(:,:,m).The function of the module is to use St',
(U1)t, (U2)t, (U3)t, Z is calculated according to following methodst'(:,:, m) and it is used as image block PtIt is filtered as a result,
Wherein：
Zt'=St'×1(U1)t×2(U2)t×3(U3)t
Wherein, ×nIndicate mould n tensor product, n=1,2 or 3；
C. Fusion Module, input include image block { Pt| t=1,2,3 ... T } obtained after above-mentioned apparatus filters
As a result { Pt' | t=1,2,3 ... T }, output is the filter result of location of pixels PP.The function of the module is to { Pt' | t=
1,2 ... T in (it,jt) pixel value of position merged.
The filter of the compression video, it is characterised in that the output image block P of the image block determining modulet
The input for finding module with similar block is connected.The similar block finds the three rank tensor Z of output of moduletMould is converted with HOSVD
The input of block is connected.The output of the HOSVD conversion module-decomposition coefficient matrix StWith the input phase of threshold process module
Even, the output τ of threshold calculation module is connected with the input of threshold process module.Output-surprise of the HOSVD conversion module
Different matrix U1, U2, U3It is connected with the input of HOSVD inverse transform block, the output S of the threshold process modulet' and HOSVD
The input of inverse transform block is connected.The output P of the HOSVD inverse transform blockt' be connected with the input of Fusion Module.
Further, the input of threshold calculation module includes quantization step Qstep, Video coding frame type, bright chroma point
Measure type, output include filtering threshold τ, described in operation include, according to input information, using one of following methods calculate
Filtering threshold τ：
● τ=aQstep+b
● τ=c (Qstep)d
Wherein coefficient value a, b, c, d is obtained according to encoding frame type, bright chroma component type, 0 < a≤0.2,0 b≤1 <,
0 < c≤1,0 d≤1 <.
Further, coefficient value a, b, c, d of chromatic component U/V are identical, different from the coefficient value of luminance component Y.
The threshold tau of luminance component is greater than the threshold tau of chromatic component；The threshold tau that the threshold tau of I frame is greater than B frame is greater than the threshold tau of P frame.
The invention adopts the above technical scheme, which has the following advantages：
Filtering algorithm based on HOSVD is a kind of simple, the filtering algorithm based on structural similarity and sparsity, the party
The method similar block three rank tensor different for content obtains different adaptive bases by training study, can preferably express
Picture material, for this particular problem of the removal of quantizing noise in compression video, using adaptively can using coding side
The method of the information export filtering threshold obtained, can more accurately estimate the degree of quantizing noise, to reach good filtering effect
Fruit.In addition, singular value decomposition needs image block being arranged in column vector, the two-dimensional structure of image block can be lost, while converting base
Using two-dimensional transform base, capable and row, column and column and the correlation between block and block are not fully taken into account, and high-order is odd
The decomposition of different value does not need for high dimensional data to be launched into matrix that the topological structure inside data will not be destroyed to decompose, so
There are more freedom degrees decomposing obtained feature space, it is thus possible to preferably utilize the relevant information between high dimensional data
And redundancy, sparse denoising is carried out in high-dimensional feature space.
Detailed description of the invention
More fully understand to have to the present invention, with reference to the accompanying drawing, to the tool of method and apparatus of the present invention
Body embodiment is described in detail, wherein：
Fig. 1 is the flow chart of filtering algorithm of the embodiment of the present invention；
Fig. 2 is the realization principle figure of the filtering method that the embodiment of the present invention proposes and device；
Fig. 3 is the n modular matrix schematic diagram of tensor resolution；
Fig. 4 is threshold process curve graph；
Fig. 5 is filter figure of the embodiment of the present invention；
Fig. 6 is the position view of the filtering method and device of the embodiment of the present invention in mixed video coding framework；
Fig. 7 is a in threshold calculations, the reference value of b.
Specific embodiment
For a further understanding of the present invention, the preferred embodiments of the invention are described below with reference to embodiment, but
It is it should be appreciated that these descriptions are only to further illustrate the features and advantages of the present invention, rather than to the claims in the present invention
Limitation.
The embodiment of the present invention one
It is the flow chart of filtering algorithm of the embodiment of the present invention shown in Fig. 1 a, Fig. 2 is the filtering side that the embodiment of the present invention proposes
The realization principle figure of method.It is specifically described below with reference to Fig. 1 a, Fig. 2.
In step s 106, it inputs as the image block P of currently pending p × q sizem(see the box of Fig. 2) and reconstruct
Image exports as three rank tensor Z, Z ∈ R of similar blockp×q×K.Image block P is used firstm, according to similarity criterion in reconstructed image
K-1 similar block similar with current block is found out, wherein similarity criterion has pixel in different realizations, such as different images block
The sum of absolute value of the difference (SAD, Sum of Absolute Differences)：
SAD=| | Pm-Pk(k≠m)||
For example in different images block pixel difference quadratic sum (SSD, Sum of Squared Differences)：
The for another example correlation (NCC, Normalized Cross Correlation) of different images block：
Wherein, PmAnd Pk(k≠m)It is two-dimensional matrix, respectively indicates image block to be processed and its a certain similar block.Certainly, ability
The technical staff in domain is appreciated that similitude matching criterior can be selected as desired there are also different forms, therefore simultaneously
It does not limit the invention.
K-1 obtained similar block and current block to be processed are formed into three rank tensor Z ∈ Rp×q×K, wherein：
Z (m, n, k)=Pk(m,n)
It is herein a kind of current block P for indicating two-dimensional matrixmA three rank tensors are arranged as with its K-1 similar block
Form puts in order, it is notable that puts in order unrelated with final filter result.
In step S108, inputs as three rank tensor Z of similar block, export as singular value matrix U1, U2, U3And decomposition coefficient
S.Higher-order Singular value decomposition process is as follows：
S=Z ×1(U1)T×2(U2)T×3(U3)T
Wherein, ×nRepresent mould n tensor product, UnIt is obtained to carry out n modular matrix singular value decomposition to the three rank tensor Z
Left singular matrix, n=1,2,3, (*)TIndicate transposition.Wherein U1Represent the correlation between image block row and row pixel, U2
Represent the correlation between column and column pixel, U3Represent the correlation in different images block between the pixel of same position.
In step S110, image block PmThere may be multiple quantization steps (Qstep)i, these quantization steps are weighted flat
(according to image block PmIn some quantization step (Qstep)iThe number of corresponding pixel), or take its median, average value
Deng deduced image block PmQuantization step be Qstep, the threshold value of filtering algorithm can be calculated by Qstep：
τ=g (Qstep)
Wherein g (*) indicates the functional relation of threshold value and Qstep.In compression video, the energy of quantization step and quantizing noise
There are certain positive correlations for amount, and when quantization step is smaller, corresponding quantizing noise is also smaller, and vice versa.
Using thresholding functions shown in Fig. 4, the value of decomposition coefficient is handled：
Wherein S (i, j, k) is the decomposition coefficient at (i, j, k), S'(i, j, k) it is hard -threshold treated corresponding position
Decomposition coefficient,It is normalized parameter.Because three rank tensors of similar image block composition are because of its interior view
As the high correlation of block, so signal component has sparsity after Higher-order Singular value decomposition, biggish decomposition coefficient represents letter
Number component, and lesser decomposition coefficient then represents noise component(s), and by adaptive threshold process, it is most of to can achieve reservation
Signal component and remove the effect of noise component(s).
In step S112, to treated, decomposition coefficient S' carries out HOSVD inverse transformation：
Z'=S' ×1(U1)×2(U2)×3(U3)
Obtain filtered three ranks tensor, Z'(:,:, m) and it is the filtered result of image block P.Wherein Z (:,:, x) and table
Show x-th of two-dimensional matrix for taking out the third dimension in three rank tensors.
The embodiment of the present invention two
It is the flow chart of filtering algorithm of the embodiment of the present invention shown in Fig. 1 b.It is specifically described below with reference to Fig. 1 b.
In step s 102, the video reconstruction image f and location of pixels PP of input are received.In step S104, middle warp is found out
Cross the image block { P of PPt| t=1,2 ... T }, wherein T is the image block number by PP, PtIn (it,jt) correspond to PP.It is right
In each image block Pt, the processing of step S106~S112 is carried out respectively.
In step s 106, it inputs as currently pending pt×qtThe image block P of sizet(see the box of Fig. 2) and reconstruct
Image exports as three rank tensor Z of similar blockt, Zt∈Rp×q×K.Use current block as template first, in reconstructed image according to
Similarity criterion finds out K-1 similar block similar with current block, and wherein similarity criterion has different realizations, such as different figures
As the sum of pixel absolute value of the difference (SAD, Sum of Absolute Differences) in block：
SAD=| | Pt-Pk(k≠t)||
For example in different images block pixel difference quadratic sum (SSD, Sum of Squared Differences)：
The for another example correlation (NCC, Normalized Cross Correlation) of different images block：
Wherein, PtAnd Pk(k≠t)It is two-dimensional matrix, respectively indicates image block to be processed and its a certain similar block.Certainly, ability
The technical staff in domain is appreciated that similitude matching criterior can be selected as desired there are also different forms, therefore simultaneously
It does not limit the invention.
K-1 obtained similar block and current block to be processed are formed into three rank tensor Z ∈ Rp×q×K, wherein：
Zt(m, n, k)=Pk(m,n)
It is herein a kind of current block P for indicating two-dimensional matrixtA three rank tensors are arranged as with its K-1 similar block
Form puts in order, it is notable that puts in order unrelated with final filter result.
In step S108, input as three rank tensor Z of similar blockt, export as singular value matrix (U1)t, (U2)t, (U3)tAnd
Decomposition coefficient St.Higher-order Singular value decomposition process is as follows：
St=Zt×1(U1)t T×2(U2)t T×3(U3)t T
Wherein, ×nRepresent mould n tensor product, (Un)tFor to the three rank tensor ZtCarry out n modular matrix singular value decomposition
Obtained left singular matrix, n=1,2,3, (*)TIndicate transposition.Wherein (U1)tRepresent the phase between image block row and row pixel
Guan Xing, U2Represent the correlation between column and column pixel, U3Represent the phase in different images block between the pixel of same position
Guan Xing.
In step S110, quantization step Qstep, the threshold value of filtering algorithm：
τ=g (Qstep)
Wherein g (*) indicates the functional relation of threshold value and Qstep.In compression video, the energy of quantization step and quantizing noise
There are certain positive correlations for amount, and when quantization step is smaller, corresponding quantizing noise is also smaller, and vice versa.
Using thresholding functions shown in Fig. 4, the value of decomposition coefficient is handled：
Wherein St(i, j, k) is the decomposition coefficient at (i, j, k), St' (i, j, k) be hard -threshold treated corresponding position
Decomposition coefficient,It is normalized parameter.Because three rank tensors of similar image block composition are because inside it
The high correlation of image block, so signal component has sparsity after Higher-order Singular value decomposition, biggish decomposition coefficient is represented
Signal component, and lesser decomposition coefficient then represents noise component(s), by adaptive threshold process, can achieve and retains big portion
Point signal component and remove the effect of noise component(s).
In step S112, to treated, decomposition coefficient S' carries out HOSVD inverse transformation：
Z'=S' ×1(U1)×2(U2)×3(U3)
Obtain filtered three ranks tensor, Z'(:,:, m) and it is the filtered result of image block P.Wherein Z (:,:, x) and table
Show x-th of two-dimensional matrix for taking out the third dimension in three rank tensors.
Image block { P in step S114 judgment step S104t| t=1,2 ... T } whether it is disposed, if not provided,
It jumps back at step S106, is disposed, carry out step S116.
In step S116, to { Pt' | t=1,2 ... T in (it,jt) pixel value of position merged, as the image
The filter result of middle position PP, fusion method can be：The pixel value for meeting above-mentioned condition is averaged, as location of pixels PP
Filtered result.The method of fusion is also possible that the median for taking all pixels value for meeting above-mentioned condition, or weighting are put down
Mean value is as result after the filtering of the location of pixels.
It will be understood to those skilled in the art that there are many kinds of the methods of fusion.Therefore, specific embodiment party given here
Formula is not construed as limiting protection scope of the present invention.Although describing a specific embodiment of the invention in conjunction with attached drawing,
Those skilled in the art can the inspiration that the technology of the present invention is conceived and on the basis of do not depart from the content of present invention to the present invention
Various deformations or amendments are made, these deformations or modification are still fallen within protection scope of the present invention.
The embodiment of the present invention three
It is the flow chart of filtering algorithm of the present invention shown in Fig. 1 b.It is specifically described below with reference to Fig. 1 b.
In step s 102, the video reconstruction image f and location of pixels PP of input are received.In step S104, middle warp is found out
Cross the image block { P of PPt| t=1,2 ... T }, wherein T is the image block number by PP, PtIn (it,jt) correspond to PP.It is right
In each image block Pt, the processing of step S106~S112 is carried out respectively.
In step s 106, it inputs as currently pending pt×qtThe image block P of sizet(see the box of Fig. 2) and reconstruct
Image exports as three rank tensor Z of similar blockt, Zt∈Rp×q×K.Use current block as template first, according to phase in reconstructed image
K-1 similar block similar with current block is found out like property criterion, wherein similarity criterion has different realizations, such as different images
The sum of pixel absolute value of the difference (SAD, Sum of Absolute Differences) in block：
SAD=| | Pt-Pk(k≠t)||
For example in different images block pixel difference quadratic sum (SSD, Sum of Squared Differences)：
The for another example correlation (NCC, Normalized Cross Correlation) of different images block：
Wherein, PtAnd Pk(k≠t)It is two-dimensional matrix, respectively indicates image block to be processed and its a certain similar block.Certainly, ability
The technical staff in domain is appreciated that similitude matching criterior can be selected as desired there are also different forms, therefore simultaneously
It does not limit the invention.
K-1 obtained similar block and current block to be processed are formed into three rank tensor Z, wherein：
Zt(m, n, k)=Pk(m,n)
It is herein a kind of current block P for indicating two-dimensional matrixtA three rank tensors are arranged as with its K-1 similar block
Form puts in order, it is notable that puts in order unrelated with final filter result.
In step S108, input as three rank tensor Z of similar blockt, export as singular value matrix (U1)t, (U2)t, (U3)tAnd
Decomposition coefficient St.Higher-order Singular value decomposition process is as follows：
St=Zt×1(U1)t T×2(U2)t T×3(U3)t T
Wherein, ×nRepresent mould n tensor product, (Un)tFor to the three rank tensor ZtCarry out n modular matrix singular value decomposition
Obtained left singular matrix, n=1,2,3, (*)TIndicate transposition.Wherein (U1)tRepresent the phase between image block row and row pixel
Guan Xing, U2Represent the correlation between column and column pixel, U3Represent the phase in different images block between the pixel of same position
Guan Xing.
In step S110, quantization step Qstep, the threshold value of filtering algorithm：
τ=g (Qstep)
Wherein g (*) indicates the functional relation of threshold value and Qstep.In compression video, the energy of quantization step and quantizing noise
There are certain positive correlations for amount, and when quantization step is smaller, corresponding quantizing noise is also smaller, and vice versa.
Using thresholding functions shown in Fig. 4, the value of decomposition coefficient is handled：
Wherein St(i, j, k) is the decomposition coefficient at (i, j, k), St' (i, j, k) be hard -threshold treated corresponding position
Decomposition coefficient,It is normalized parameter.Because three rank tensors of similar image block composition are because inside it
The high correlation of image block, so signal component has sparsity after Higher-order Singular value decomposition, biggish decomposition coefficient is represented
Signal component, and lesser decomposition coefficient then represents noise component(s), by adaptive threshold process, can achieve and retains big portion
Point signal component and remove the effect of noise component(s).
In step S112, to treated, decomposition coefficient S' carries out HOSVD inverse transformation：
Z'=S' ×1(U1)×2(U2)×3(U3)
Obtain filtered three ranks tensor, Z'(:,:, m) and it is the filtered result of image block P.Wherein Z (:,:, x) and table
Show x-th of two-dimensional matrix for taking out the third dimension in three rank tensors.
Image block { P in step S114 judgment step S104t| t=1,2 ... T } whether it is disposed, if not provided,
It jumps back at step S106, is disposed, carry out step S116.
In step S116, to { Pt' | t=1,2 ... T in (it,jt) pixel value of position merged, as the image
The filter result of middle position PP, fusion method can be：The pixel value for meeting above-mentioned condition is averaged, as location of pixels PP
Filtered result.The method of fusion is also possible that the median for taking all pixels value for meeting above-mentioned condition, or weighting are put down
Mean value is as result after the filtering of the location of pixels.
This filtering only changes output image, is added without in coding and decoding video loop as post-processing operation.
The embodiment of the present invention four
It is the flow chart of filtering algorithm of the present invention shown in Fig. 1 b.It is specifically described below with reference to Fig. 1 b.
In step s 102, the video reconstruction image f and location of pixels PP of input are received.In step S104, middle warp is found out
Cross the image block { P of PPt| t=1,2 ... T }, wherein T is the image block number by PP, PtIn (it,jt) correspond to PP.It is right
In each image block Pt, the processing of step S106~S112 is carried out respectively.
In step s 106, it inputs as currently pending pt×qtThe image block P of sizet(see the box of Fig. 2) and reconstruct
Image exports as three rank tensor Z of similar blockt, Zt∈Rp×q×K.Use current block as template first, according to phase in reconstructed image
K-1 similar block similar with current block is found out like property criterion, wherein similarity criterion has different realizations, such as different images
The sum of pixel absolute value of the difference (SAD, Sum of Absolute Differences) in block：
SAD=| | Pt-Pk(k≠t)||
For example in different images block pixel difference quadratic sum (SSD, Sum of Squared Differences)：
The for another example correlation (NCC, Normalized Cross Correlation) of different images block：
Wherein, PtAnd Pk(k≠t)It is two-dimensional matrix, respectively indicates image block to be processed and its a certain similar block.Certainly, ability
The technical staff in domain is appreciated that similitude matching criterior can be selected as desired there are also different forms, therefore simultaneously
It does not limit the invention.
K-1 obtained similar block and current block to be processed are formed into three rank tensor Z, wherein：
Zt(m, n, k)=Pk(m,n)
It is herein a kind of current block P for indicating two-dimensional matrixtA three rank tensors are arranged as with its K-1 similar block
Form puts in order, it is notable that puts in order unrelated with final filter result.
In step S108, input as three rank tensor Z of similar blockt, export as singular value matrix (U1)t, (U2)t, (U3)tAnd
Decomposition coefficient St.Higher-order Singular value decomposition process is as follows：
St=Zt×1(U1)t T×2(U2)t T×3(U3)t T
Wherein, ×nRepresent mould n tensor product, (Un)tFor to the three rank tensor ZtCarry out n modular matrix singular value decomposition
Obtained left singular matrix, n=1,2,3, (*)TIndicate transposition.Wherein (U1)tRepresent the phase between image block row and row pixel
Guan Xing, U2Represent the correlation between column and column pixel, U3Represent the phase in different images block between the pixel of same position
Guan Xing.
In step S110, quantization step Qstep, the threshold value of filtering algorithm：
τ=g (Qstep)
Wherein g (*) indicates the functional relation of threshold value and Qstep.In compression video, the energy of quantization step and quantizing noise
There are certain positive correlations for amount, and when quantization step is smaller, corresponding quantizing noise is also smaller, and vice versa.
Using thresholding functions shown in Fig. 4, the value of decomposition coefficient is handled：
Wherein St(i, j, k) is the decomposition coefficient at (i, j, k), St' (i, j, k) be hard -threshold treated corresponding position
Decomposition coefficient,It is normalized parameter.Because three rank tensors of similar image block composition are because inside it
The high correlation of image block, so signal component has sparsity after Higher-order Singular value decomposition, biggish decomposition coefficient is represented
Signal component, and lesser decomposition coefficient then represents noise component(s), by adaptive threshold process, can achieve and retains big portion
Point signal component and remove the effect of noise component(s).
In step S112, to treated, decomposition coefficient S' carries out HOSVD inverse transformation：
Z'=S' ×1(U1)×2(U2)×3(U3)
Obtain filtered three ranks tensor, Z'(:,:, m) and it is the filtered result of image block P.Wherein Z (:,:, x) and table
Show x-th of two-dimensional matrix for taking out the third dimension in three rank tensors.
Image block { P in step S114 judgment step S104t| t=1,2 ... T } whether it is disposed, if not provided,
It jumps back at step S106, is disposed, carry out step S116.
In step S116, to { Pt' | t=1,2 ... T in (it,jt) pixel value of position merged, as the image
The filter result of middle position PP, fusion method can be：The pixel value for meeting above-mentioned condition is averaged, as location of pixels PP
Filtered result.The method of fusion is also possible that the median for taking all pixels value for meeting above-mentioned condition, or weighting are put down
Mean value is as result after the filtering of the location of pixels.
This filtering method is added in coding and decoding video loop, filtering image is put into reference picture caching, as
The reference picture of inter-prediction further increases code efficiency so as to improve the performance of inter-prediction.
The embodiment of the present invention five
It is the flow chart of filtering algorithm of the present invention shown in Fig. 1 (b), Fig. 6 is that the filtering method that the embodiment of the present invention proposes is answered
Position in mixed video frame.It is specifically described below with reference to Fig. 1 (b), Fig. 6.
The filter is applied into the loop filter position in hybrid video coding, filter proposed by the present invention
Video coding loop inverse quantization reconstructing video can be located at and compress any position between video output caching, as shown in fig. 6,
The present embodiment and the embodiment of the present invention 4 the difference is that, 3. the filter of embodiment 4 is applied to be located in position, in addition to this,
It can be applied in and 1. or 2. locate.
In step s 102, the video reconstruction image f and location of pixels PP of input are received.In step S104, middle warp is found out
Cross the image block { P of PPt| t=1,2 ... T }, wherein T is the image block number by PP, PtIn (it,jt) correspond to PP.It is right
In each image block Pt, the processing of step S106~S112 is carried out respectively.
In step s 106, it inputs as currently pending pt×qtThe image block P of sizet(see the box of Fig. 2) and reconstruct
Image exports as three rank tensor Z of similar blockt, Zt∈Rp×q×K.Use current block as template first, according to phase in reconstructed image
K-1 similar block similar with current block is found out like property criterion, wherein similarity criterion has different realizations, such as different images
The sum of pixel absolute value of the difference (SAD, Sum of Absolute Differences) in block：
SAD=| | Pt-Pk(k≠t)||
For example in different images block pixel difference quadratic sum (SSD, Sum of Squared Differences)：
The for another example correlation (NCC, Normalized Cross Correlation) of different images block：
Wherein, PtAnd Pk(k≠t)It is two-dimensional matrix, respectively indicates image block to be processed and its a certain similar block.Certainly, ability
The technical staff in domain is appreciated that similitude matching criterior can be selected as desired there are also different forms, therefore simultaneously
It does not limit the invention.
K-1 obtained similar block and current block to be processed are formed into three rank tensor Z ∈ Rp×q×K, wherein：
Zt(m, n, k)=Pk(m,n)
It is herein a kind of current block P for indicating two-dimensional matrixtA three rank tensors are arranged as with its K-1 similar block
Form puts in order, it is notable that puts in order unrelated with final filter result.
In step S108, input as three rank tensor Z of similar blockt, export as singular value matrix (U1)t, (U2)t, (U3)tAnd
Decomposition coefficient St.Higher-order Singular value decomposition process is as follows：
St=Zt×1(U1)t T×2(U2)t T×3(U3)t T
Wherein, ×nRepresent mould n tensor product, (Un)tFor to the three rank tensor ZtCarry out n modular matrix singular value decomposition
Obtained left singular matrix, n=1,2,3, (*)TIndicate transposition.Wherein (U1)tRepresent the phase between image block row and row pixel
Guan Xing, U2Represent the correlation between column and column pixel, U3Represent the phase in different images block between the pixel of same position
Guan Xing.
In step S110, quantization step Qstep, the threshold value of filtering algorithm：
τ=g (Qstep)
Wherein g (*) indicates the functional relation of threshold value and Qstep.In compression video, the energy of quantization step and quantizing noise
There are certain positive correlations for amount, and when quantization step is smaller, corresponding quantizing noise is also smaller, and vice versa.
Using thresholding functions shown in Fig. 4, the value of decomposition coefficient is handled：
Wherein St(i, j, k) is the decomposition coefficient at (i, j, k), St' (i, j, k) be hard -threshold treated corresponding position
Decomposition coefficient,It is normalized parameter.Because three rank tensors of similar image block composition are because inside it
The high correlation of image block, so signal component has sparsity after Higher-order Singular value decomposition, biggish decomposition coefficient is represented
Signal component, and lesser decomposition coefficient then represents noise component(s), by adaptive threshold process, can achieve and retains big portion
Point signal component and remove the effect of noise component(s).
In step S112, to treated, decomposition coefficient S' carries out HOSVD inverse transformation：
Z'=S' ×1(U1)×2(U2)×3(U3)
Obtain filtered three ranks tensor, Z'(:,:, m) and it is the filtered result of image block P.Wherein Z (:,:, x) and table
Show x-th of two-dimensional matrix for taking out the third dimension in three rank tensors.
Image block { P in step S114 judgment step S104t| t=1,2 ... T } whether it is disposed, if not provided,
It jumps back at step S106, is disposed, carry out step S116.
In step S116, to { Pt' | t=1,2 ... T in (it,jt) pixel value of position merged, as the image
The filter result of middle position PP, fusion method can be：The pixel value for meeting above-mentioned condition is averaged, as location of pixels PP
Filtered result.The method of fusion is also possible that the median for taking all pixels value for meeting above-mentioned condition, or weighting are put down
Mean value is as result after the filtering of the location of pixels.
Aforesaid operations are made to each pixel in image, finally obtain filtering image.Filtering image is exported, as coding
The input of next stage in loop.
The embodiment of the present invention six
A kind of filtering side based on non local spatial correlation and Higher-order Singular value decomposition (HOSVD) for compression video
Method, it is same as Example 1 that it includes steps, and difference is：Device block to be filtered is searched in reconstructed frame using the method for template matching
When rope similar block, it is not fixed the number of similar block, but according to matching criterior fixing search threshold value, meet the block i.e. conduct of condition
Similar block.
The embodiment of the present invention seven
A kind of filtering side based on non local spatial correlation and Higher-order Singular value decomposition (HOSVD) for compression video
Method, it is same as Example 1 that it includes steps, and difference is：The selection of threshold value is not only related with quantization step Qstep, also with view
The type of coding of frequency frame, YUV component type is related, and the relationship of fitting is：
τ=aQstep+b
Or
τ=c (Qstep)d
Wherein coefficient value a, b, c, d is obtained according to encoding frame type, bright chroma component type, 0 < a≤0.2,0 b≤1 <,
0 < c≤1,0 d≤1 <.Wherein, the more excellent value of a, b, c, d are shown in Fig. 7
The embodiment of the present invention eight
It is filter apparatus figure of the present invention shown in Fig. 5 (a), is specifically described below with reference to Fig. 5 (a).Module S202,
It inputs image block P and video reconstruction image f including p × q size to be processed, and output includes three rank tensor Z of similar block,
Z∈Rp×q×K.The function of the module is, with image block P, the method in f using Block- matching finds K-1 similar block, wherein phase
There can be the sum of pixel absolute value of the difference (SAD, Sum of in different realizations, such as different images block like property criterion
Absolute Differences)：
SAD=| | Pm-Pk(k≠m)||
For example in different images block pixel difference quadratic sum (SSD, Sum of Squared Differences)：
The for another example correlation (NCC, Normalized Cross Correlation) of different images block：
Wherein, PmAnd Pk(k≠m)It is two-dimensional matrix, respectively indicates image block to be processed and its a certain similar block.Certainly, ability
The technical staff in domain is appreciated that similitude matching criterior can be selected as desired there are also different forms, therefore simultaneously
It does not limit the invention.
K-1 obtained similar block and image block P are formed into three rank tensor Z ∈ Rp×q×K, wherein：
Z (m, n, k)=Pk(m,n)
It is herein a kind of current block P for indicating two-dimensional matrixmA three rank tensors are arranged as with its K-1 similar block
Form puts in order, it is notable that puts in order unrelated with final filter result.The output of module S200 is current
Image block PmWith the three rank tensor Z of p × q × K size of its K-1 similar block composition.
Module S204, input include three rank tensor Z of similar block, and output includes the orthogonal singular matrix U of three units1,
U2, U3With decomposition coefficient matrix S ∈ R.The function of the module is to carry out higher order singular value point to the three rank tensor Z of similar block
Solution, obtains singular matrix U1, U2, U3With decomposition coefficient S.
Meanwhile module S206 carries out threshold calculations, input includes quantization step Qstep, and output includes filtering threshold
τ.The function of the module is, according to the quantization step Qstep, to calculate the threshold tau of filtering algorithm.
τ=g (Qstep)
Module S208, output τ and the HOSVD conversion module of the input including the threshold calculation module are defeated
Decomposition coefficient S out, output include the decomposition coefficient S' after threshold process.The function of the module is, to HOSVD conversion module
The decomposition coefficient S of output carries out following threshold process：
Wherein St(i, j, k) is the decomposition coefficient at (i, j, k), St' (i, j, k) be hard -threshold treated corresponding position
Decomposition coefficient,It is normalized parameter.Because three rank tensors of similar image block composition are because inside it
The high correlation of image block, so signal component has sparsity after Higher-order Singular value decomposition, biggish decomposition coefficient is represented
Signal component, and lesser decomposition coefficient then represents noise component(s), by adaptive threshold process, can achieve and retains big portion
Point signal component and remove the effect of noise component(s).
In module S210, input includes three singular matrix U of HOSVD conversion module output1, U2, U3And threshold
It is worth processing module output treated decomposition coefficient S', output includes the filtered result Z'(of image block P:,:,m).It should
The function of module is to utilize S', U1, U2, U3, calculate filtered three ranks tensor Z'：
Z'=S' ×1(U1)×2(U2)×3(U3)
Wherein, ×nIndicate mould n tensor product, n=1,2 or 3.Z'(:,:, m) and it is the filtered result of image block P.Wherein
Z(:,:, x) and it indicates to take out x-th of two-dimensional matrix of the third dimension in three rank tensors.
The embodiment of the present invention nine
It is filter apparatus figure of the present invention shown in Fig. 5 (b), below with reference to Fig. 5, (b0 is specifically described.
Module S200, input include three rank tensor Z of similar block for its output of video reconstruction image f and location of pixels PP,
Z∈Rp×q×K.The function of the module is that, with image block P, the method in f using Block- matching finds K-1 similar block；By image
Block P forms three rank tensor Z together with K-1 similar block；Wherein Z (:,:, m) and it is image block P.
Module S202, input include p to be processedt×qtThe image block P of sizetWith video reconstruction image f, output
Including three rank tensor Z of similar block.It includes three rank tensor Z, Z ∈ R of similar block that it, which is exported,p×q×K.The function of the module is to use image
Block Pt, the method in f using Block- matching finds K-1 similar block, and wherein similarity criterion can have different realizations, such as
The sum of pixel absolute value of the difference (SAD, Sum of Absolute Differences) in different images block：
SAD=| | Pt-Pk(k≠t)||
For example in different images block pixel difference quadratic sum (SSD, Sum of Squared Differences)：
The for another example correlation (NCC, Normalized Cross Correlation) of different images block：
Wherein, PtAnd Pk(k≠t)It is two-dimensional matrix, respectively indicates image block to be processed and its a certain similar block.Certainly, ability
The technical staff in domain is appreciated that similitude matching criterior can be selected as desired there are also different forms, therefore simultaneously
It does not limit the invention.
K-1 obtained similar block and current block to be processed are formed into three rank tensor Z ∈ Rp×q×K, wherein：
Zt(m, n, k)=Pk(m,n)
It is herein a kind of current block P for indicating two-dimensional matrixtA three rank tensors are arranged as with its K-1 similar block
Form puts in order, it is notable that puts in order unrelated with final filter result.The output of module S200 is current
Image block PtWith the three rank tensor Z of p × q × K size of its K-1 similar block compositiont。
Module S204, input include three rank tensor Z of similar blockt, output includes the orthogonal singular matrix of three units
(U1)t, (U2)t, (U3)tAnd decomposition coefficient matrix St.The function of the module is, to the three rank tensor Z of similar blocktIt carries out
Higher-order Singular value decomposition.
Meanwhile module S206 carries out threshold calculations, input includes quantization step Qstep, and output includes filtering threshold
τ.The function of the module is, according to the quantization step Qstep, to calculate the threshold tau of filtering algorithm.
τ=g (Qstep)
Module S208, output τ and the HOSVD conversion module of the input including the threshold calculation module are defeated
Decomposition coefficient S outt, output includes the decomposition coefficient S after threshold processt'.The function of the module is to convert mould to HOSVD
The decomposition coefficient S of block outputtCarry out following threshold process：
Wherein St(i, j, k) is the decomposition coefficient at (i, j, k), St' (i, j, k) be hard -threshold treated corresponding position
Decomposition coefficient,It is normalized parameter.Because three rank tensors of similar image block composition are because inside it
The high correlation of image block, so signal component has sparsity after Higher-order Singular value decomposition, biggish decomposition coefficient is represented
Signal component, and lesser decomposition coefficient then represents noise component(s), by adaptive threshold process, can achieve and retains big portion
Point signal component and remove the effect of noise component(s).
In module S210, input includes three singular matrix U of HOSVD conversion module output1, U2, U3And threshold
It is worth processing module output treated decomposition coefficient S', output includes the filtered result Z'(of image block P:,:,m).It should
The function of module is to utilize S', U1, U2, U3, calculate filtered three ranks tensor Z'：
Z'=S' ×1(U1)×2(U2)×3(U3)
Wherein, ×nIndicate mould n tensor product, n=1,2 or 3.Z'(:,:, m) and it is the filtered result of image block P.Wherein
Z(:,:, x) and it indicates to take out x-th of two-dimensional matrix of the third dimension in three rank tensors.
To { Pt| t=1,2 ... T } in all image blocks execute aforesaid operations after, to { Pt' | t=1,2 ... T } in
(it,jt) pixel value of position merged, as the filter result of the position PP in the image, fusion method be can be：It will meet
The pixel value of above-mentioned condition is averaged, as the filtered result of location of pixels PP.The method of fusion, which is also possible that, to be taken in satisfaction
The median or weighted average for stating all pixels value of condition are as result after the filtering of the location of pixels.
The embodiment of the present invention ten
It is filter apparatus figure of the present invention shown in Fig. 5 (b), is specifically described below with reference to Fig. 5 (b).
Module S200, input include three rank tensor Z of similar block for its output of video reconstruction image f and location of pixels PP,
Z∈Rp×q×K.The function of the module is that, with image block P, the method in f using Block- matching finds K-1 similar block；By image
Block P forms three rank tensor Z together with K-1 similar block；Wherein Z (:,:, m) and it is image block P.
Module S202, input include p to be processedt×qtThe image block P of sizetWith video reconstruction image f, output
Including three rank tensor Z of similar block.It includes three rank tensor Z, Z ∈ R of similar block that it, which is exported,p×q×K.The function of the module is to use image
Block Pt, the method in f using Block- matching finds K-1 similar block, and wherein similarity criterion can have different realizations, such as
The sum of pixel absolute value of the difference (SAD, Sum of Absolute Differences) in different images block：
SAD=| | Pt-Pk(k≠t)||
For example in different images block pixel difference quadratic sum (SSD, Sum of Squared Differences)：
The for another example correlation (NCC, Normalized Cross Correlation) of different images block：
Wherein, PtAnd Pk(k≠t)It is two-dimensional matrix, respectively indicates image block to be processed and its a certain similar block.Certainly, ability
The technical staff in domain is appreciated that similitude matching criterior can be selected as desired there are also different forms, therefore simultaneously
It does not limit the invention.
K-1 obtained similar block and current block to be processed are formed into three rank tensor Z ∈ Rp×q×K, wherein：
Zt(m, n, k)=Pk(m,n)
It is herein a kind of current block P for indicating two-dimensional matrixtA three rank tensors are arranged as with its K-1 similar block
Form puts in order, it is notable that puts in order unrelated with final filter result.The output of module S200 is current
Image block PtWith the three rank tensor Z of p × q × K size of its K-1 similar block compositiont。
Module S204, input include three rank tensor Z of similar blockt, output includes the orthogonal singular matrix of three units
(U1)t, (U2)t, (U3)tAnd decomposition coefficient matrix St.The function of the module is, to the three rank tensor Z of similar blocktIt carries out
Higher-order Singular value decomposition.
Meanwhile module S206 carries out threshold calculations, input includes quantization step Qstep, and output includes filtering threshold
τ.The function of the module is, according to the quantization step Qstep, to calculate the threshold tau of filtering algorithm.
τ=g (Qstep)
Module S208, output τ and the HOSVD conversion module of the input including the threshold calculation module are defeated
Decomposition coefficient S outt, output includes the decomposition coefficient S after threshold processt'.The function of the module is to convert mould to HOSVD
The decomposition coefficient S of block outputtCarry out following threshold process：
Wherein St(i, j, k) is the decomposition coefficient at (i, j, k), St' (i, j, k) be hard -threshold treated corresponding position
Decomposition coefficient,It is normalized parameter.Because three rank tensors of similar image block composition are because inside it
The high correlation of image block, so signal component has sparsity after Higher-order Singular value decomposition, biggish decomposition coefficient is represented
Signal component, and lesser decomposition coefficient then represents noise component(s), by adaptive threshold process, can achieve and retains big portion
Point signal component and remove the effect of noise component(s).
In module S210, input includes three singular matrix U of HOSVD conversion module output1, U2, U3And threshold
It is worth processing module output treated decomposition coefficient S', output includes the filtered result Z'(of image block P:,:,m).It should
The function of module is to utilize S', U1, U2, U3, calculate filtered three ranks tensor Z'：
Z'=S' ×1(U1)×2(U2)×3(U3)
Wherein, ×nIndicate mould n tensor product, n=1,2 or 3.Z'(:,:, m) and it is the filtered result of image block P.Wherein
Z(:,:, x) and it indicates to take out x-th of two-dimensional matrix of the third dimension in three rank tensors.
To { Pt| t=1,2 ... T } in all image blocks execute aforesaid operations after, to { Pt' | t=1,2 ... T } in
(it,jt) pixel value of position merged, as the filter result of the position PP in the image, fusion method be can be：It will meet
The pixel value of above-mentioned condition is averaged, as the filtered result of location of pixels PP.The method of fusion, which is also possible that, to be taken in satisfaction
The median or weighted average for stating all pixels value of condition are as result after the filtering of the location of pixels.
This device is used to post-process compression video, above-mentioned processing is carried out to the pixel in compressed video frame, is only changed
Become output image, the image in reference picture caching is not changed.
The embodiment of the present invention 11
Fig. 6 (is filter apparatus figure of the present invention shown in b0, is specifically described below with reference to Fig. 6 (b).
Module S200, input include three rank tensor Z of similar block for its output of video reconstruction image f and location of pixels PP,
Z∈Rp×q×K.The function of the module is that, with image block P, the method in f using Block- matching finds K-1 similar block；By image
Block P forms three rank tensor Z together with K-1 similar block；Wherein Z (:,:, m) and it is image block P.
Module S202, input include p to be processedt×qtThe image block P of sizetWith video reconstruction image f, output
Including three rank tensor Z of similar block.It includes three rank tensor Z, Z ∈ R of similar block that it, which is exported,p×q×K.The function of the module is to use image
Block Pt, the method in f using Block- matching finds K-1 similar block, and wherein similarity criterion can have different realizations, such as
The sum of pixel absolute value of the difference (SAD, Sum of Absolute Differences) in different images block：
SAD=| | Pt-Pk(k≠t)||
For example in different images block pixel difference quadratic sum (SSD, Sum of Squared Differences)：
The for another example correlation (NCC, Normalized Cross Correlation) of different images block：
Wherein, PtAnd Pk(k≠t)It is two-dimensional matrix, respectively indicates image block to be processed and its a certain similar block.Certainly, ability
The technical staff in domain is appreciated that similitude matching criterior can be selected as desired there are also different forms, therefore simultaneously
It does not limit the invention.
K-1 obtained similar block and current block to be processed are formed into three rank tensor Z, wherein：
Zt(m, n, k)=Pk(m,n)
It is herein a kind of current block P for indicating two-dimensional matrixtA three rank tensors are arranged as with its K-1 similar block
Form puts in order, it is notable that puts in order unrelated with final filter result.The output of module S200 is current
Image block PtWith the three rank tensor Z of p × q × K size of its K-1 similar block compositiont。
Module S204, input include three rank tensor Z of similar blockt, output includes the orthogonal singular matrix of three units
(U1)t, (U2)t, (U3)tAnd decomposition coefficient matrix St.The function of the module is, to the three rank tensor Z of similar blocktIt carries out
Higher-order Singular value decomposition obtains singular matrix (U1)t, (U2)t, (U3)tAnd decomposition coefficient matrix St。
Meanwhile module S206 carries out threshold calculations, input includes quantization step Qstep, and output includes filtering threshold
τ.The function of the module is, according to the quantization step Qstep, to calculate the threshold tau of filtering algorithm.
τ=g (Qstep)
Module S208, output τ and the HOSVD conversion module of the input including the threshold calculation module are defeated
Decomposition coefficient S outt, output includes the decomposition coefficient S after threshold processt'.The function of the module is to convert mould to HOSVD
The decomposition coefficient S of block outputtCarry out following threshold process：
Wherein St(i, j, k) is the decomposition coefficient at (i, j, k), St' (i, j, k) be hard -threshold treated corresponding position
Decomposition coefficient,It is normalized parameter.Because three rank tensors of similar image block composition are because inside it
The high correlation of image block, so signal component has sparsity after Higher-order Singular value decomposition, biggish decomposition coefficient is represented
Signal component, and lesser decomposition coefficient then represents noise component(s), by adaptive threshold process, can achieve and retains big portion
Point signal component and remove the effect of noise component(s).
In module S210, input includes three singular matrix U of HOSVD conversion module output1, U2, U3And threshold
It is worth processing module output treated decomposition coefficient S', output includes the filtered result Z'(of image block P:,:,m).It should
The function of module is to utilize S', U1, U2, U3, calculate filtered three ranks tensor Z'：
Z'=S' ×1(U1)×2(U2)×3(U3)
Wherein, ×nIndicate mould n tensor product, n=1,2 or 3.Z'(:,:, m) and it is the filtered result of image block P.Wherein
Z(:,:, x) and it indicates to take out x-th of two-dimensional matrix of the third dimension in three rank tensors.
To { Pt| t=1,2 ... T } in all image blocks execute aforesaid operations after, to { Pt' | t=1,2 ... T } in
(it,jt) the corresponding pixel value in position merged, as the filter result of the position PP in the image, fusion method be can be：It will
The pixel value for meeting above-mentioned condition is averaged, as the filtered result of location of pixels PP.The method of fusion, which is also possible that, to be taken completely
The median or weighted average of all pixels value of sufficient above-mentioned condition are as result after the filtering of the location of pixels.
Above-mentioned apparatus is added in encoding and decoding loop device, filtering image is put into reference picture caching, as interframe
The reference picture of prediction further increases code efficiency so as to improve the performance of inter-prediction
The embodiment of the present invention 12
It is filter apparatus figure of the present invention shown in Fig. 5 (b), Fig. 6 is that the filter that the embodiment of the present invention proposes is applied
Position in mixed video frame.It is specifically described below with reference to Fig. 5 (b), Fig. 6.
The filter apparatus is applied into the loop filter position in hybrid video coding, filter proposed by the present invention
Wave device device can be located at Video coding loop inverse quantization reconstructing video and compress any position between video output caching, such as
Shown in Fig. 6, the present embodiment and the embodiment of the present invention 9 the difference is that, 3. the filter of embodiment 9 is applied to be located in position,
Locate in addition to this it is possible to apply 1. or 2..
Module S200, input include three rank tensor Z of similar block for its output of video reconstruction image f and location of pixels PP,
Z∈Rp×q×K.The function of the module is that, with image block P, the method in f using Block- matching finds K-1 similar block；By image
Block P forms three rank tensor Z together with K-1 similar block；Wherein Z (:,:, m) and it is image block P.
Module S202, input include p to be processedt×qtThe image block P of sizetWith video reconstruction image f, output
Including three rank tensor Z of similar block.It includes three rank tensor Z, Z ∈ R of similar block that it, which is exported,p×q×K.The function of the module is to use image
Block Pt, the method in f using Block- matching finds K-1 similar block, and wherein similarity criterion can have different realizations, such as
The sum of pixel absolute value of the difference (SAD, Sum of Absolute Differences) in different images block：
SAD=| | Pt-Pk(k≠t)||
For example in different images block pixel difference quadratic sum (SSD, Sum of Squared Differences)：
The for another example correlation (NCC, Normalized Cross Correlation) of different images block：
Wherein, PtAnd Pk(k≠t)It is two-dimensional matrix, respectively indicates image block to be processed and its a certain similar block.Certainly, ability
The technical staff in domain is appreciated that similitude matching criterior can be selected as desired there are also different forms, therefore simultaneously
It does not limit the invention.
K-1 obtained similar block and current block to be processed are formed into three rank tensor Z ∈ Rp×q×K, wherein：
Zt(m, n, k)=Pk(m,n)
It is herein a kind of current block P for indicating two-dimensional matrixtA three rank tensors are arranged as with its K-1 similar block
Form puts in order, it is notable that puts in order unrelated with final filter result.The output of module S200 is current
Image block PtWith the three rank tensor Z of p × q × K size of its K-1 similar block compositiont。
Module S204, input include three rank tensor Z of similar blockt, output includes the orthogonal singular matrix of three units
(U1)t, (U2)t, (U3)tAnd decomposition coefficient matrix St.The function of the module is, to the three rank tensor Z of similar blocktIt carries out
Higher-order Singular value decomposition.
Meanwhile module S206 carries out threshold calculations, input includes quantization step Qstep, and output includes filtering threshold
τ.The function of the module is, according to the quantization step Qstep, to calculate the threshold tau of filtering algorithm.
τ=g (Qstep)
Module S208, output τ and the HOSVD conversion module of the input including the threshold calculation module are defeated
Decomposition coefficient S outt, output includes the decomposition coefficient S after threshold processt'.The function of the module is to convert mould to HOSVD
The decomposition coefficient S of block outputtCarry out following threshold process：
Wherein St(i, j, k) is the decomposition coefficient at (i, j, k), St' (i, j, k) be hard -threshold treated corresponding position
Decomposition coefficient,It is normalized parameter.Because three rank tensors of similar image block composition are because inside it
The high correlation of image block, so signal component has sparsity after Higher-order Singular value decomposition, biggish decomposition coefficient is represented
Signal component, and lesser decomposition coefficient then represents noise component(s), by adaptive threshold process, can achieve and retains big portion
Point signal component and remove the effect of noise component(s).
In module S210, input includes three singular matrix U of HOSVD conversion module output1, U2, U3And threshold
It is worth processing module output treated decomposition coefficient S', output includes the filtered result Z'(of image block P:,:,m).It should
The function of module is to utilize S', U1, U2, U3, calculate filtered three ranks tensor Z'：
Z'=S' ×1(U1)×2(U2)×3(U3)
Wherein, ×nIndicate mould n tensor product, n=1,2 or 3.Z'(:,:, m) and it is the filtered result of image block P.Wherein
Z(:,:, x) and it indicates to take out x-th of two-dimensional matrix of the third dimension in three rank tensors.
To { Pt| t=1,2 ... T } in all image blocks execute aforesaid operations after, to { Pt' | t=1,2 ... T } in
(it,jt) pixel value of position merged, as the filter result of the position PP in the image, fusion method be can be：It will meet
The pixel value of above-mentioned condition is averaged, as the filtered result of location of pixels PP.The method of fusion, which is also possible that, to be taken in satisfaction
The median or weighted average for stating all pixels value of condition are as result after the filtering of the location of pixels.
It is handled to obtain filtering image using above-mentioned filter apparatus.Filtering image is exported, as in coding loop
The input of next stage.
The embodiment of the present invention 13
A kind of filtering dress based on non local spatial correlation and Higher-order Singular value decomposition (HOSVD) for compression video
It sets, it is same as Example 8 that it includes modules, and difference is：As shown in Fig. 6 (c), the selection of threshold value not only with quantization step
Qstep is related, and also with the type of coding of video frame, YUV component type is related, and the relationship of fitting is one of following：
τ=aQstep+b
τ=c (Qstep)d
Wherein coefficient value a, b, c, d is obtained according to encoding frame type, bright chroma component type, 0 < a≤0.2,0 b≤1 <,
0 < c≤1,0 d≤1 <.Wherein, the more excellent value of a, b, c, d are shown in Fig. 7.
Claims (11)
1. a kind of filtering method for compressing video, including：
A.
● using the image block P of p × q size to be processed, K-1 similar block is found using Block- matching in reconstructed image；It will figure
As block P forms three rank tensor Z, Z ∈ R together with K-1 similar blockp×q×K；Wherein Z (:,:, m) and it is image block P；
● according to the quantization step Qstep of image block P, determine the threshold tau of filtering algorithm：
τ=g (Qstep)
B. Higher-order Singular value decomposition is carried out to the three rank tensor Z, obtains singular value matrix U1, U2, U3And decomposition coefficient square
Battle array S, wherein U1∈Rp×p, U2∈Rq×q, U3∈RK×K, S ∈ Rp×q×K；
C. according to the filtering threshold τ, each of S element is handled using threshold method, obtains S'：
Wherein S (i, j, k) is the resolving system numerical value in decomposition coefficient matrix S at (i, j, k), S'(i, j, k) it is at S (i, j, k)
After reason in S' corresponding position value；
D. with decomposition coefficient the matrix S', singular value matrix U after threshold process1, U2, U3, Z' is calculated according to following methods
(:,:, m) and filtered as a result, wherein as image block P：
Z'=S' ×1(U1)×2(U2)×3(U3)
×nIndicate mould n tensor product, n=1,2 or 3.
2. the filtering method of compression video according to claim 1, it is characterised in that：
The g (Qstep) is following one：
● g (Qstep)=aQstep+b
● g (Qstep)=c (Qstep)d
Wherein coefficient value a, b, c, d is obtained according to encoding frame type, bright chroma component type, 0 < a≤0.2,0 < b≤1,0 < c
≤ 1,0 d≤1 <.
3. the filtering method of compression video according to claim 2, it is characterised in that：
Coefficient value a, b, c, d of chromatic component U/V is identical, different from the coefficient value of luminance component Y；The threshold of the luminance component
Value τ is greater than the threshold tau of chromatic component, and the threshold tau that the threshold tau of I frame is greater than B frame is greater than the threshold tau of P frame.
4. a kind of filtering method for compressing video, it is characterised in that：
A. for a certain location of pixels PP in image f, the middle image block { P by PP is found outt| t=1,2 ... T }, wherein T is
By the image block number of PP, PtIn (it,jt) correspond to PP；
B. for each image block Pt,It is handled by the following method：
a)
● utilize p to be processedt×qtThe image block P of sizet, K-1 similar block is found using Block- matching in reconstructed image；It will
Image block PtThree rank tensor Z are formed together with K-1 similar blockt, Zt∈Rp×q×K；Wherein Zt(:,:, m) and it is image block Pt；
● according to image block PtQuantization step Qstep, determine the threshold tau of filtering algorithm：
τ=g (Qstep)
B) to the three rank tensor ZtHigher-order Singular value decomposition is carried out, singular value matrix (U is obtained1)t, (U2)t, (U3)tAnd point
Solve coefficient matrix St, wherein (U1)t∈Rp×p, (U2)t∈Rq×q, (U3)t∈RK×K, St∈Rp×q×K；
C) according to the filtering threshold τ, using threshold method to StEach of element handled, obtain S 't：
Wherein St(i, j, k) is decomposition coefficient matrix StIn resolving system numerical value at (i, j, k), St' (i, j, k) be St(i,j,k)
S ' after processingtThe value of middle corresponding position；
D) the decomposition coefficient matrix S after threshold process is utilizedt', singular value matrix (U1)t, (U2)t, (U3)tAccording to following methods meter
Calculation obtains Zt'(:,:, m) and it is used as image block PtIt is filtered as a result, wherein：
Zt'=St'×1(U1)t×2(U2)t×3(U3)t
×nIndicate mould n tensor product, n=1,2 or 3；
C. to { Pt' | t=1,2 ... T in (it,jt) pixel value of position merged, the filter as position PP in the image
Wave result.
5. the filtering method of compression video according to claim 4, it is characterised in that：
The g (Qstep) is following one：
● g (Qstep)=aQstep+b
● g (Qstep)=c (Qstep)d
Wherein coefficient value a, b, c, d is obtained according to encoding frame type, bright chroma component type, 0 < a≤0.2,0 < b≤1,0 < c
≤ 1,0 d≤1 <.
6. the filtering method of compression video according to claim 5, it is characterised in that：The coefficient value a of chromatic component U/V,
B, c, d are identical, different from the coefficient value of luminance component Y；The threshold tau of the luminance component is greater than the threshold tau of chromatic component, I
The threshold tau that the threshold tau of frame is greater than B frame is greater than the threshold tau of P frame.
7. a kind of filter for compression video, it is characterised in that including：
A.
● similar block finds module, and input includes the image block P and video reconstruction image f of p × q size to be processed, defeated
It out include three rank tensor Z, Z ∈ R of similar blockp×q×K, the similar block finds module and is used for：With image block P, block is utilized in f
The method matched finds K-1 similar block；Image block P is formed to three rank tensor Z together with K-1 similar block；Wherein Z (:,:,m)
For image block P；
● threshold calculation module, input include quantization step Qstep, and output includes filtering threshold τ；The threshold calculations mould
Block is used to calculate the threshold tau of filtering algorithm according to the quantization step Qstep；
B.HOSVD conversion module, input include three rank tensor Z of similar block, and output includes singular matrix U1, U2, U3And point
Coefficient matrix S is solved, wherein U1∈Rp×p, U2∈Rq×q, U3∈RK×K, S ∈ Rp×q×K；The HOSVD conversion module is used for described
Three rank tensor Z of similar block carry out Higher-order Singular value decomposition；
C. threshold process module, input include the threshold tau and the decomposition coefficient matrix S, and output includes threshold
Value treated decomposition coefficient matrix S'.The function of the module is to carry out to each of decomposition coefficient matrix S element as follows
Threshold process obtains S'：
Wherein S (i, j, k) is the resolving system numerical value in decomposition coefficient matrix S at (i, j, k), S'(i, j, k) it is at S (i, j, k)
Corresponding position value in S' after reason；
D.HOSVD inverse transform block, input include singular matrix U1, U2, U3, and treated decomposition coefficient matrix S',
Output includes the filtered result Z'(of image block P:,:, m), the HOSVD inverse transform block is used for, with S', U1, U2, U3, root
Z'(is calculated according to following methods:,:, m) and filtered as a result, wherein as image block P：
Z'=S' ×1(U1)×2(U2)×3(U3)
Wherein, ×nIndicate mould n tensor product, n=1,2 or 3；
The filter of the compression video, it is characterised in that also reside in, the similar block finds the output-tri- of module
Rank tensor Z is connected with the input of HOSVD conversion module, the output-decomposition coefficient matrix S and threshold of the HOSVD conversion module
The input for being worth processing module is connected, and output-threshold tau of threshold calculation module is connected with threshold process module, the HOSVD
The output of conversion module-singular matrix U1, U2, U3Be connected with the input of HOSVD inverse transform block, threshold process module it is defeated
Out-treated, and decomposition coefficient matrix S' is connected with the input of HOSVD inverse transform block.
8. the filter according to claim 7 for compression video, which is characterized in that
The input of threshold calculation module includes quantization step Qstep, Video coding frame type, bright chroma component type, output
Including filtering threshold, described in operation include, according to input information, calculating filtering threshold τ：
● τ=aQstep+b
● τ=c (Qstep)d
Wherein coefficient value a, b, c, d is obtained according to encoding frame type, bright chroma component type, 0 < a≤0.2,0 < b≤1,0 < c
≤ 1,0 d≤1 <.
9. the filter according to claim 8 for compression video, which is characterized in that in threshold calculation module：
Coefficient value a, b, c, d of chromatic component U/V is identical, different from the coefficient value of luminance component Y；The threshold of the luminance component
Value τ is greater than the threshold tau of chromatic component, and the threshold tau that the threshold tau of I frame is greater than B frame is greater than the threshold tau of P frame.
10. a kind of filter for compression video, it is characterised in that including：
A. image block determining module, input include a certain location of pixels PP in video reconstruction image f and f, output packet
It includes by the image block { P by PPt| t=1,2 ... T }, wherein T is the image block number by PP, PtIn (it,jt) right
It answers；Described image block determining module is used to find out all image blocks for passing through location of pixels PP in image f；
B. for each image block Pt,It is handled using following apparatus：
(a)
● similar block finds module, and input includes p to be processedt×qtThe image block P of sizetWith video reconstruction image f,
Output includes three rank tensor Z of similar blockt, Zt∈Rp×q×KThe similar block is found module and is used for：With image block Pt, utilized in f
The method of Block- matching finds K-1 similar block；By image block PtThree rank tensor Z are formed together with K-1 similar blockt；Wherein Zt
(:,:, m) and it is image block Pt；
● threshold calculation module, input include quantization step Qstep, and output includes filtering threshold τ；The threshold calculations mould
Block is used to calculate the threshold tau of filtering algorithm according to the quantization step Qstep；
(b) HOSVD conversion module, input include three rank tensor Z of similar blockt, output includes singular matrix (U1)t, (U2)t,
(U3)tAnd decomposition coefficient matrix St, wherein (U1)t∈Rp×p, (U2)t∈Rq×q, (U3)t∈RK×K, St∈Rp×q×K；It is described
HOSVD conversion module is used for the three rank tensor Z of similar blocktCarry out Higher-order Singular value decomposition；
(c) threshold process module, input include the threshold tau and the decomposition coefficient matrix St, output include
Decomposition coefficient matrix S ' after threshold processt, the threshold process module is for decomposition coefficient matrix StEach of element
Following threshold process is carried out, S ' is obtainedt：
Wherein St(i, j, k) is decomposition coefficient matrix StIn resolving system numerical value at (i, j, k), St' (i, j, k) be S (i, j, k)
S ' after processingtMiddle corresponding position value；
(d) HOSVD inverse transform block, input include singular matrix (U1)t, (U2)t, (U3)t, and treated decomposition coefficient
Matrix St', output includes image block PtFiltered result Zt'(:,:,m)；The HOSVD inverse transform block is used for, and is used
St', (U1)t, (U2)t, (U3)t, Z is calculated according to following methodst'(:,:, m) and it is used as image block PtIt is filtered as a result, its
In：
Zt'=St'×1(U1)t×2(U2)t×3(U3)t
Wherein, ×nIndicate mould n tensor product, n=1,2 or 3；
C. Fusion Module, input include image block { Pt| t=1,2,3 ... T } result that is obtained after above-mentioned apparatus filters
{Pt' | t=1,2,3 ... T }, output is the filter result of location of pixels PP；The Fusion Module is used for as to { Pt' | t=
1,2 ... T in (it,jt) pixel value of position merged；
The filter of the compression video, it is further characterized in that, the output image block P of the image block determining moduletWith
The input that similar block finds module is connected, and the similar block finds the three rank tensor Z of output of moduletWith HOSVD conversion module
Input be connected, output-decomposition coefficient matrix S of the HOSVD conversion moduletIt is connected with the input of threshold process module,
The output τ of threshold calculation module is connected with the input of threshold process module, the unusual square of output-of the HOSVD conversion module
Battle array U1, U2, U3It is connected with the input of HOSVD inverse transform block, the output S of the threshold process modulet' and HOSVD inverse transformation
The input of module is connected, the output P of the HOSVD inverse transform blockt' be connected with the input of Fusion Module.
11. the filter according to claim 10 for compression video, which is characterized in that
The input of threshold calculation module includes quantization step Qstep, Video coding frame type, bright chroma component type, output
Including filtering threshold, described in operation include, according to input information, calculating filtering threshold τ：
● τ=aQstep+b
● τ=c (Qstep)d
Wherein coefficient value a, b, c, d is obtained according to encoding frame type, bright chroma component type, 0 < a≤0.2,0 < b≤1,0 < c
≤ 1,0 d≤1 <.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710386993.4A CN108933942B (en) | 2017-05-27 | 2017-05-27 | Filtering method of compressed video and filtering device for compressed video |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710386993.4A CN108933942B (en) | 2017-05-27 | 2017-05-27 | Filtering method of compressed video and filtering device for compressed video |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108933942A true CN108933942A (en) | 2018-12-04 |
| CN108933942B CN108933942B (en) | 2020-10-30 |
Family
ID=64451590
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710386993.4A Active CN108933942B (en) | 2017-05-27 | 2017-05-27 | Filtering method of compressed video and filtering device for compressed video |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108933942B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110446067A (en) * | 2019-08-30 | 2019-11-12 | 杭州电子科技大学 | Video concentration method based on tensor resolution |
| CN114302150A (en) * | 2021-12-30 | 2022-04-08 | 北京超维景生物科技有限公司 | Video encoding method and device, video decoding method and device, and electronic device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110292216A1 (en) * | 2009-01-09 | 2011-12-01 | New York University | Method, computer-accessible, medium and systems for facilitating dark flash photography |
| CN103093434A (en) * | 2013-01-27 | 2013-05-08 | 西安电子科技大学 | Non-local wiener filtering image denoising method based on singular value decomposition |
| CN104200441A (en) * | 2014-09-18 | 2014-12-10 | 南方医科大学 | Higher-order singular value decomposition based magnetic resonance image denoising method |
| CN105763881A (en) * | 2016-02-23 | 2016-07-13 | 北京大学 | Video coding image filtering method based on image nonlocal structural similarity characteristic |
-
2017
- 2017-05-27 CN CN201710386993.4A patent/CN108933942B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110292216A1 (en) * | 2009-01-09 | 2011-12-01 | New York University | Method, computer-accessible, medium and systems for facilitating dark flash photography |
| CN103093434A (en) * | 2013-01-27 | 2013-05-08 | 西安电子科技大学 | Non-local wiener filtering image denoising method based on singular value decomposition |
| CN104200441A (en) * | 2014-09-18 | 2014-12-10 | 南方医科大学 | Higher-order singular value decomposition based magnetic resonance image denoising method |
| CN105763881A (en) * | 2016-02-23 | 2016-07-13 | 北京大学 | Video coding image filtering method based on image nonlocal structural similarity characteristic |
Non-Patent Citations (2)
| Title |
|---|
| VLADIMIR KATKOVNIK,ET AL.: "Sparse phase imaging based on complex domain nonlocal BM3D techniques", 《DIGITAL SIGNAL PROCESSING》 * |
| 张鑫媛: "基于结构相似性的磁共振图像去噪新算法研究", 《中国博士学位论文全文数据库》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110446067A (en) * | 2019-08-30 | 2019-11-12 | 杭州电子科技大学 | Video concentration method based on tensor resolution |
| CN110446067B (en) * | 2019-08-30 | 2021-11-02 | 杭州电子科技大学 | Tensor decomposition-based video concentration method |
| CN114302150A (en) * | 2021-12-30 | 2022-04-08 | 北京超维景生物科技有限公司 | Video encoding method and device, video decoding method and device, and electronic device |
| CN114302150B (en) * | 2021-12-30 | 2024-02-27 | 北京超维景生物科技有限公司 | Video encoding method and device, video decoding method and device and electronic equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108933942B (en) | 2020-10-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP7110370B2 (en) | Using Nonlinear Functions Applied to Quantization Parameters of Machine Learning Models for Video Coding | |
| CN107743239B (en) | Method and device for encoding and decoding video data | |
| CN107197260A (en) | Video coding post-filter method based on convolutional neural networks | |
| EP1577790A2 (en) | Integer transform matrix selection method in video coding and related integer transform matrix | |
| CN107645662B (en) | Color image compression method | |
| Narmatha et al. | A lossless compression scheme for grayscale medical images using a P2-bit short technique | |
| CN106658004B (en) | A kind of compression method and device based on image flat site feature | |
| CN110073663A (en) | Use the transformation coefficient code of rank figure | |
| CN110392256A (en) | Method, coding side, decoding end and the system coded and decoded | |
| Zhao et al. | Image compressive-sensing recovery using structured laplacian sparsity in DCT domain and multi-hypothesis prediction | |
| CN109196861B (en) | Method for decoding digital image, encoding method and device | |
| CN108933942A (en) | A kind of filtering method compressing video and the filter for compression video | |
| PraveenKumar et al. | Medical image compression using integer multi wavelets transform for telemedicine applications | |
| CN110352599A (en) | Method for video processing and equipment | |
| CN109819258A (en) | A kind of splits' positions perceived direction predictive coding method based on spiral scan | |
| CN109819251B (en) | Encoding and decoding method of pulse array signal | |
| Naidu et al. | A novel framework for JPEG image compression using baseline coding with parallel process | |
| Awad et al. | Improving Reconstructed Image Quality via Hybrid Compression Techniques. | |
| Peter Devadoss et al. | Region of interest based MRI brain image compression using Peano space filling curve | |
| Kumar et al. | Comparative Analysis and Performance Evaluation of Medical Image Compression Method for Telemedicine | |
| JP2003304537A (en) | Method of processing digital image for low-bit rate application | |
| Maalouf et al. | Low-complexity enhanced lapped transform for image coding in JPEG XR/HD photo | |
| CN110175965A (en) | Based on splits' positions cognitive method that is adaptively sampled and smoothly projecting | |
| Nihal et al. | A Survey and Study of Image Compression Methods | |
| Pawar et al. | A review: Approximate image and video compression |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |