Embodiment
Fig. 1 is the flow chart of video image denoising embodiment of the method one of the present invention, and as shown in Figure 1, the method comprises:
Step 101, employing preset Denoising Algorithm, carry out predenoising process, obtain reference video image to pending first video image;
Step 102, for the pending image block in current frame image in described first video image, in described reference video image, determine that the image block corresponding with described pending image block has the reference image block group of similitude, described reference image block group comprises the similar image block in the two field picture corresponding with described current frame image, and the similar image block in each two field picture in certain frame number adjacent with described current frame image;
Step 103, the coordinate position relation of described reference image block group is mapped in described first video image from described reference video image, obtain image block group similar to described pending image block in the first video image, and low-rank approximation process is carried out to the described image block group in described first video image, obtain the second video image of denoising.
In the present embodiment, suppose that the first video image of pending band noise is Y, and Y=X+N, wherein, X represents the clean video image in this first video image, and N is noise.Namely the described method that the present embodiment provides is to remove noise N, obtains clean video image X.
First, adopt and preset Denoising Algorithm, predenoising process is carried out to pending first video image Y, obtains reference video image
wherein, the Denoising Algorithm with better performance of above-mentioned default Denoising Algorithm such as usually adopting in prior art, such as based on the Denoising Algorithm, V-BM4D algorithm etc. of wavelet transformation.In the present embodiment, predenoising process is carried out to obtain reference video image to the first video image Y
object be, at reference video image
in determine and current frame image Y in the first video image Y
kin any pending image block y
i,kcorresponding image block
there is each reference image block of structural similarity
Particularly, for current frame image Y in the first video image Y
kin pending image block y
i,k, it is at reference video image
the image block of middle correspondence is
namely have and image block y
i,kidentical position coordinates i, and then, in reference video image, find with
there is each reference image block of structural similarity
wherein, the measurement criterion of this similitude such as with the Euclidean distance of two image blocks for foundation, or with the F norm distance of two image blocks for foundation, not to be limited with the present embodiment.And, in the present embodiment, due to the continuity of the different inter frame images of video image, determine with
there is each reference image block of structural similarity
process in, be not only at reference video image
in with the present frame Y in the first video image Y
kcorresponding two field picture
in search, can also at this corresponding two field picture
search in multiple frames within adjacent certain frame number, such as
in, namely this reference image block group comprises and described current frame image Y
kcorresponding two field picture
in similar image block, and with described current frame image Y
ksimilar image block in each two field picture in adjacent certain frame number.
In the present embodiment, the coordinate set searching each reference image block determined is designated as Ω
i,k, and remember that the reference image block group that each reference image block is formed is
be understandable that,
also be with current frame image Y in the first video image Y
kin image block y
i,kthere is each reference image block of structural similarity
the similarity matrix of composition.
And then, with reference to image block group
coordinate position relation be mapped to the first video image from reference video image, thus obtain in the first video image with pending image block y
i,ksimilar image block group
and then to image block group each in described first video image
carry out low-rank approximation process, obtain the second video image of denoising.Wherein, namely described low-rank approximation process is to the image block group in the first video image
carry out singular value decomposition, obtain the low-rank matrix corresponding to singular value being greater than certain threshold value, this singular value threshold method can adopt successive ignition mode to determine this image block group rank of matrix, with this low-rank matrix approximate matrix
and then again singular value decomposition inverse transformation and weighted average are carried out to obtain the second video image of denoising to low-rank matrix
wherein, above-mentioned threshold value both can be the hard-threshold of fixed numbers, also can be the soft-threshold of a threshold function table, was preferably the latter.
Optionally, above-mentioned denoising process can iteration repeatedly, to improve denoising effect further.
The present invention carries in video image Denoising Algorithm exists outer circulation and Inner eycle two iterative process, and wherein outer circulation iteration upgrades reference video image, and Inner eycle iteration upgrades the first video image be with and made an uproar.Particularly, for being with first video image of making an uproar, Inner eycle can using the iteration relative error of iterations or adjacent twice as the condition of convergence.In iterative process, suppose that initialization operation is:
Y
(0)=Y,
In solution procedure, Inner eycle iteration more new formula is:
wherein, u is step-length, and t is current iteration number of times.
In the present embodiment, adopt and preset Denoising Algorithm, first predenoising process is carried out to pending first video image, obtain reference video image, and then in described reference video image, determine that the image block corresponding with arbitrary pending image block in current frame image in the first video image has the reference image block group of similitude, the coordinate position relation of this reference image block group is mapped to the first video image from reference video image, thus obtain image block group similar to pending image block in the first video image, and then low-rank approximation process is carried out to image block group similar to pending image block in the first video image, obtain the second video image of denoising.Using the foundation of the reference video image arbitrary picture block structure similarity measurement in the first video image through predenoising process, because the noise existed in reference video image is relatively less, make similarity measurement result more accurate, and be conducive to retaining the detailed information such as edge, texture in video image by this similarity measurement, thus be conducive to the quality of the second video image ensureing final denoising; In addition, the mode of approaching based on low-rank solves the second video image obtaining final denoising from the first video image, further avoid the adverse effect that in the weighted average of the overall situation in non-local mean mode, single-phase produces video image denoising effect with weights, improve the denoising effect of video image.
Fig. 2 is the flow chart of video image denoising embodiment of the method two of the present invention, and as shown in Figure 2, the method comprises:
Step 201, according to formula (7), described first video image is converted to YCbCr form, the form of described first video image is rgb format:
Form due to most of video image of reality is rgb format, and rgb format has stronger correlation, be unfavorable for denoising, therefore, first the first video image of rgb format is converted to FR YCbCr form in the present embodiment, particularly, the matrix relationship of formula (7) is adopted to change.Thus carry out following denoising for vectorial Y, Cb and Cr respectively, result is stacked together the most at last, obtain the video image of denoising.
Step 202, employing preset Denoising Algorithm, carry out predenoising process, obtain reference video image to pending first video image;
First, adopt and preset Denoising Algorithm, predenoising process is carried out to pending first video image Y, obtains reference video image
Step 203, in described reference video image, determine that the image block corresponding with described pending image block has the reference image block group of similitude according to formula (1)
Wherein,
be in described reference video image with the described pending image block y in the current frame image in described first video image
i,kcorresponding image block, described y
i,kin comprise
individual pixel, i is the center point coordinate position of described pending image block, and k is the frame number of described present frame,
for arbitrary and described in described reference video image
f norm distance be less than the reference image block of little positive number ε, j is the coordinate position of reference image block, and the frame number of the frame of l belonging to reference image block, l equals or is not equal to k, Ω
i,kfor all coordinate sets meeting the reference image block of formula (1).
Described in step 204, general
be mapped to from described reference video image in described first video image, obtain image block group similar to described pending image block in the first video image
and according to formula (2) to the image block group in described first video image
carry out low-rank approximation process:
Wherein, || ||
w, *for Weighted Kernel norm, λ is regular parameter,
for unknown variable,
Wherein,
for matrix
a jth singular value,
For weight coefficient matrix,
Wherein, c is constant, and n is the described image block group in the first video image
in the number of image block that comprises, ξ avoids the little positive number divided by 0, singular value
initial estimate be:
Wherein, σ
nit is the noise standard deviation of the first video image.
Wherein,
for the optimal value on the right of equation, determine according to formula (3):
Wherein, S
w() is respectively right for soft-threshold function, matrix Σ, U and V
carry out singular value coefficient matrix and the unitary matrice of singular value decomposition gained:
wherein, singular value coefficient matrix Σ=diag{ α
1, α
2..., α
p, α
j, j=1,2..., P are singular value, according to S
w(Σ) the jth singular value α determined
jthreshold value be:
Step 205, to determine according to formula (4) described in
order r:
r=sum(j=1,…,P|α
j>τ
j)(4)
During specific implementation, in basis
obtain ω
jafter, calculate a jth singular value α
jcorresponding threshold tau
j, thus obtain
order r, from formula (4), if α
j> τ
j, then α
jretain, otherwise, α
jbe placed in 0.
Step 206, obtain the second video image of denoising according to formula (5):
Wherein, weights
determine according to formula (6):
From above-mentioned formula (5), obtaining the second pure video image of denoising as wanted, needing the estimated value determining image block group in the second video image
with this estimated value
weighting function.Wherein, the estimated value of image block group
can according to formula (3), brought the estimated value asking for image block group in the second video image that low-rank represents by singular value decomposition inversion, obtain weights corresponding to image block group estimated value according to formula (6).And then obtain pure second video image according to weighted average formula (5)
And then with described second video image for current iteration inputted video image, upgrade described reference video image, outer circulation iteration performs following steps 207-208:
Step 207, in current iteration inputted video image, determine that the image block corresponding with described pending image block has another reference image block group of similitude, and another reference image block group described is mapped in described first video image from described current iteration inputted video image, obtain another image block group similar to described pending image block in the first video image; Low-rank approximation process is carried out to another image block group similar to described pending image block in described first video image, obtains the following iteration inputted video image of denoising;
Step 208, determine whether described following iteration inputted video image meets iteration cut-off condition, if meet, then perform step 209, otherwise perform step 207;
Wherein, described iteration cut-off condition comprises: iterations reaches preset times, or the relative error of the video image that adjacent iteration exports is less than predetermined threshold value.
Further, denoising process in the present embodiment is the process that a successive ignition performs, namely obtain second video image of denoising of an iterative process acquisition according to step 202-206 after, the foundation of alternative reference video image as the structural similarity image block of the image block determined in the first video image is upgraded using this second video image, the image block group similar to any image block in current frame image in the first video image is determined in the first video image, and then low-rank approximation process is carried out to image block group each in the first video image, the second video image obtaining denoising upgrades, i.e. the 3rd video image, and then using the 3rd video image as current iteration inputted video image, upgrade reference video image, iteration performs, until meet above-mentioned iteration cut-off condition, final video image is exported.
Step 209, export described following iteration inputted video image as final denoising video image;
Step 210, according to formula (8), described second video image is converted to rgb format:
Accordingly, when the first video image of rgb format is converted to YCbCr form by step 201, after obtaining the final denoising video image exported, the video image after this denoising is converted to rgb format again, adopts formula (8) to realize.
In the present embodiment, first predenoising process is carried out to pending first video image, obtain reference video image, and then in described reference video image, determine that the image block corresponding with any pending image block in current frame image in the first video image has the reference image block group of structural similarity, the coordinate position relation of this similar reference image block group is mapped to the first video image from reference picture, thus obtain image block group similar to described pending image block in the first video image, and then low-rank approximation process is carried out to this image block group in the first video image, obtain the second video image of denoising.Using the foundation of the reference video image arbitrary picture block structure similarity measurement in the first video image through predenoising process, because the noise existed in reference video image is relatively less, make similarity measurement result more accurate, and be conducive to retaining the detailed information such as edge, texture in video image by this similarity measurement, thus be conducive to the quality of the second video image ensureing final denoising; In addition, the mode of approaching based on low-rank solves the second video image obtaining final denoising from the first video image, further avoid the adverse effect that in the weighted average of the overall situation in non-local mean mode, single-phase produces video image denoising effect with weights, improve the denoising effect of video image; In addition, successive ignition execution denoising process also makes final denoising effect more optimize; And, denoising video image is set according to varying in size of the order solved there are different weighting weights in above-mentioned formula (6), make denoising effect more obvious.
Fig. 3 is the schematic diagram of video image denoising device embodiment one of the present invention, and as shown in Figure 3, this device comprises:
First processing module 11, for adopting default Denoising Algorithm, carrying out predenoising process to pending first video image, obtaining reference video image;
Determination module 12, for for the pending image block in current frame image in described first video image, in described reference video image, determine that the image block corresponding with described pending image block has the reference image block group of similitude, described reference image block group comprises the similar image block in the two field picture corresponding with described current frame image, and the similar image block in each two field picture in certain frame number adjacent with described current frame image;
Second processing module 13, for the coordinate position relation of described reference image block group is mapped in described first video image from described reference video image, obtain image block group similar to described pending image block in the first video image, and low-rank approximation process is carried out to the described image block group in described first video image, obtain the second video image of denoising.
The device of the present embodiment may be used for the technical scheme performing embodiment of the method shown in Fig. 1, and it realizes principle and technique effect is similar, repeats no more herein.
Fig. 4 is the schematic diagram of video image denoising device embodiment two of the present invention, as shown in Figure 4, the device that the present embodiment provides on basis embodiment illustrated in fig. 3, described determination module 12, specifically for:
In described reference video image, determine that the image block corresponding with described pending image block has the reference image block group of similitude according to formula (1)
Wherein,
be in described reference video image with the described pending image block y in the current frame image in described first video image
i,kcorresponding image block, described y
i,kin comprise
individual pixel, i is the center point coordinate position of described pending image block, and k is the frame number of described present frame,
for arbitrary and described in described reference video image
f norm distance be less than the reference image block of little positive number ε, j is the coordinate position of reference image block, and the frame number of the frame of l belonging to reference image block, l equals or is not equal to k, Ω
i,kfor all coordinate sets meeting the reference image block of formula (1);
Accordingly, described second processing module 13, described in inciting somebody to action
be mapped to from described reference video image in described first video image, obtain image block group similar to described pending image block in the first video image
Further, described second processing module 13, comprising:
First processing unit 131, according to formula (2) to the image block group in described first video image
carry out low-rank approximation process:
Wherein, || ||
w, *for Weighted Kernel norm, λ is regular parameter,
for unknown variable,
Wherein,
for matrix
a jth singular value,
For weight coefficient matrix,
Wherein, c is constant, and n is the described image block group in the first video image
in the number of image block that comprises, ξ avoids the little positive number divided by 0, singular value
initial estimate be:
Wherein, σ
nit is the noise standard deviation of the first video image;
Wherein,
for the optimal value on the right of equation, determine according to formula (3):
Wherein, S
w() is respectively right for soft-threshold function, matrix Σ, U and V
carry out singular value coefficient matrix and the unitary matrice of singular value decomposition gained:
wherein, singular value coefficient matrix Σ=diag{ α
1, α
2..., α
p, α
j, j=1,2..., P are singular value, according to S
w(Σ) the jth singular value α determined
jthreshold value be:
Second processing unit 132, described in determining according to formula (4)
order r:
r=sum(j=1,…,P|α
j>τ
j)(4)
3rd processing unit 133, for obtaining the second video image of denoising according to formula (5):
Wherein, weights
determine according to formula (6):
Further, the form of described first video image is rgb format, and described device, also comprises:
First modular converter 21, for described first video image being converted to YCbCr form according to formula (7):
Second modular converter 22, for described second video image being converted to rgb format according to formula (8):
Particularly, described determination module 12 specifically for described second video image for current iteration inputted video image, upgrade described reference video image, and iteration perform:
In current iteration inputted video image, determine that the image block corresponding with described pending image block has another reference image block group of similitude, and another reference image block group described is mapped in described first video image from described current iteration inputted video image, obtain another image block group similar to described pending image block in the first video image;
Described second processing module 13 performs specifically for iteration:
Low-rank approximation process is carried out to another image block group similar to described pending image block in described first video image, obtains the following iteration inputted video image of denoising;
Described device also comprises:
Judge module 23, for determining whether described following iteration inputted video image meets iteration cut-off condition, if meet, then exports described following iteration inputted video image as final denoising video image;
Wherein, described outer circulation iteration cut-off condition comprises: iterations reaches preset times, or the relative error of the video image that adjacent iteration exports is less than predetermined threshold value.
The device of the present embodiment may be used for the technical scheme performing embodiment of the method shown in Fig. 2, and it realizes principle and technique effect is similar, repeats no more herein.
One of ordinary skill in the art will appreciate that: all or part of step realizing said method embodiment can have been come by the hardware that program command is relevant, aforesaid program can be stored in a computer read/write memory medium, this program, when performing, performs the step comprising said method embodiment; And aforesaid storage medium comprises: ROM, RAM, magnetic disc or CD etc. various can be program code stored medium.
Last it is noted that above each embodiment is only in order to illustrate technical scheme of the present invention, be not intended to limit; Although with reference to foregoing embodiments to invention has been detailed description, those of ordinary skill in the art is to be understood that: it still can be modified to the technical scheme described in foregoing embodiments, or carries out equivalent replacement to wherein some or all of technical characteristic; And these amendments or replacement, do not make the essence of appropriate technical solution depart from the scope of various embodiments of the present invention technical scheme.