A kind of mixing Poisson, the video denoising method of gaussian sum impulsive noise
Technical field:
Present invention relates generally to image and video information process field, refer in particular to a kind of mixing Poisson, gaussian sum pulse and make an uproar
The video denoising method of sound.
Background technology:
Although digital camera and digital sensor are developed rapidly in recent years, in acquisition of information and transmitting procedure
In, video data is often subject to the influence of a variety of noises.Noise in vision signal can cause the sense of discomfort in visual effect, and
And [1] such as the performance of some follow-up video processnig algorithms, including feature extraction, target detection and motion trackings can be influenceed.Cause
This, video denoising is still an active research direction, and with network cameras and mobile phone camera a large amount of uses and body
Reveal prior meaning.In order to recover raw information, many scholar's designs from noise pollution and the video data degraded
Various video Denoising Algorithm is used for solving this problem [2-4].Due to the high acquisition rate of digital camera, the quality of video data
Often below static single image data, its quality problems include low signal-to-noise ratio and low resolution etc..Meanwhile, video data phase
Than having higher space-time redundancy for static image data, more rich scene letter can be provided than static image data
Breath.Therefore to the restoration algorithm effect for the video that degrades, can depend greatly on this space-time redundancy by effective profit
With, and any degree can be excavated.
In the imaging system of reality, in addition to it can produce noise similar to the internal system of white Gaussian noise, also
Poisson and impulsive noise are there is, wherein poisson noise embodies the fluctuation [5] in photon counting, and impulsive noise then embodies sensing
The collection of device signal and transmission error.Conventional video Denoising Algorithm mainly handles single type noise, for mixed noise denoising
Effect is often undesirable.Although the processing of Poisson, Gauss and pulse mixed noise is significant in actual environment,
Research in terms of such mixed noise video denoising algorithm is still very limited.
The present invention provides a kind of mixing Poisson, the video denoising method of gaussian sum impulsive noise, can be more reasonably to this
Plant mixed noise to be modeled and handle, can be obtained than existing method more preferably denoising effect for this kind of mixed noise,
There is important theory and practice meaning for the video denoising under Poisson, gaussian sum pulse mixed noise environment.
[1]A.Bovik,Handbook of Image and Video Processing,Academic Press,San
Diego,CA,2000.
[2]K.Dabov,A.Foi,and K.Egiazarian,Video denoising by sparse 3-D
transform-domain collaborative filtering,IEEE trans.Eur.Signal Process.Conf.,
Poznan,Poland,1257-1260,Sep.2007.
[3]Maggioni M,Boracchi G,Foi A,et al.Video denoising,deblocking,and
enhancement through separable 4-d nonlocal spatiotemporal transforms.IEEE
Transactions on image processing,2012,21(9):3952-3966.
[4]Hui Ji,Sibin Huang,Zuowei Shen,and Yuhong Xu,Robust Video
Restoration by Joint Sparse and Low Rank Matrix Approximation,SIAM J.IMAGING
SCIENCES,4(4),pp.1122-1142,2011.
[5]D.L.Snyder,A.M.Hammoud,and R.L.White.Image recovery from data
acquired with a charge-coupled-device camera.J.Opt.Soc.Am.A,10(1983),1014-
1023.
[6]Z.Shen,K.-C.Toh,and S.Yun,An accelerated proximal gradient
algorithm for frame-based image restoration via the balanced approach,SIAM
J.Imaging Sci.,4(2011),573-596.
The content of the invention:
The present invention for conventional video Denoising Algorithm mainly handle single type noise and for mixed noise denoising effect
Undesirable the problem of, a kind of mixing Poisson, the video denoising method of gaussian sum impulsive noise are disclosed.
In order to solve the above technical problems, technical scheme proposed by the present invention is:
A kind of mixing Poisson, the video denoising method of gaussian sum impulsive noise, it is characterised in that:
Step one:The each two field picture of noise video is divided into image block, Block- matching is passed through in noise sequence of frames of video
Method is to each image block search similar image block.The each two field picture of noise video is divided into the specific method of image block
It is:
For noise sequence of frames of video { I1,I2,…,In(n be noise video frame number), by each two field picture be divided into s ×
Overlapping region width is w between the image block of s pixel sizes, image blockoIndividual pixel.For each image block P, pass through block
Method of completing the square searches for similar image block in noise sequence of frames of video, and obtains m similar image block in each frame;For each
Its all row is in turn connected to form a column vector, then the column vector collection of all similar blocks is combined into by similar image block:{di}(i
=1,2 ..., mn).On this basis, observation block matrix is constituted with all similar block column vectorsWherein D's is each
Row are all a similar block column vector, Rh×wIt is the real number matrix set that h, columns are w to represent line number.
Image block matching method in the step one, is including but not limited to entered based on Euclidean distance between image block vector
The method of row image Block- matching;
Step 2:Construct Poisson-Gaussian mixed noise priori;Its concrete form is:
Wherein, F (A, E, Z) represents Poisson-Gaussian mixed noise priori;A is the block matrix that clear block column vector is constituted, A
Each be classified as the corresponding clear block column vector of foregoing similar block;E represents block-matching error block matrix, and each of E is classified as foregoing
The corresponding matching error column vector of similar block;Z represents impulsive noise matrix, and each of Z is classified as the corresponding pulse of foregoing similar block
Noise vector;D represents the observation block matrix that similar block column vector is constituted;Max (X) representing matrixs X greatest member, division arithmetic
It is that matrix is carried out by element with square root calculation;σ is mixes in Poisson, gaussian sum poisson noise, the standard of Gaussian noise part
Difference.And N all elements are all 1.
Step 3:The video denoising model that construction is set up under mixing Poisson, gaussian sum impulse noise environment;Concrete form
For:
Solve denoising block matrix A, the error block matrix E and impulsive noise matrix Z for causing target function value to reach minimum:
Wherein, A is the block matrix that clear block column vector is constituted, and E represents block-matching error block matrix, and Z represents impulsive noise
Matrix, F (A, E, Z) represents Poisson-Gaussian mixed noise priori, | | | |*, | | | |1With | | | |FRespectively matrix core model
Number, l1Norm and Frobenius norms, λ, μ and η are constant coefficient.After solving picture rich in detail block matrix A, A row to
Amount is remapped as s × s image blocks, obtains the corresponding picture rich in detail block of similar image block;
Step 4:Denoising model is solved using optimized algorithm, the picture rich in detail block after denoising is obtained;Solve picture rich in detail block
Optimized algorithm, include but are not limited to existing algorithm Accelerated Proximal Gradient (APG).
Step 5:Picture rich in detail merged block after denoising is generated into clear video image.Its specific method is:
For kth frame (k=1,2 ..., n) clear video image, if including pixel pkThe picture rich in detail set of blocks of (x, y)
For { b1,b2,…,bv}.Wherein, x, y are respectively line number and columns of the pixel in video image, pk(x, y) is (x, y) place
Pixel, v is includes pixel pkThe image block number of (x, y), { b1,b2,…,bvRepresent to include pixel pk(x's, y) is clear
Image block set.Then pixel pkThe gray value of (x, y) is:
Wherein, I (pk(x, y)) it is pixel pkThe gray value of (x, y),For bj(j=1,2 ..., v) in pixel
pk(x, y) corresponding gray value, v is to include pixel pkThe image block number of (x, y).
Beneficial effects of the present invention:(1) Poisson, the situation of three kinds of noise mixing of gaussian sum pulse are considered, and reasonably
Mixed model is modeled.Certain type of noise or a small amount of mixed noise are only considered in the existing method that compares, the present invention
Model with close to real video noise situation.(2) present invention proposes the video denoising method for three kinds of mixed noises, phase
Existing mixing denoising method is contrasted, final denoising effect of the invention is more preferably obvious, and denoising video quality is higher.
Brief description of the drawings:
Fig. 1 is the flow chart of the method for the invention;
Fig. 2 is the present invention and the visual effect comparison diagram of other video denoising methods;In Fig. 2, the 1st row is clear noiseless
Video image;2nd row is the noise video image added after Poisson, gaussian sum impulsive noise;3rd row is VBM3D algorithms [2]
Denoising effect;4th row is VBM4D algorithms [3] denoising effect;5th row is RPCA algorithms [4] denoising effect;6th row is this method
Denoising effect;
In the case of Fig. 3 is different noise level and experiment test videos, the present invention and the PSNR of other video denoising methods
Index comparing result.
Embodiment:
Below in conjunction with the accompanying drawings 1, the present invention will be described in detail:
A kind of mixing Poisson of the present embodiment offer, the video denoising method of gaussian sum impulsive noise, comprise the following steps:
Step one:The each two field picture of noise video is divided into image block, Block- matching is passed through in noise sequence of frames of video
Method is to each image block search similar image block.The each two field picture of noise video is divided into the specific method of image block
It is:
For noise sequence of frames of video { I1,I2,…,In(n be noise video frame number), by each two field picture be divided into s ×
Overlapping region width is w between the image block of s pixel sizes, image blockoIndividual pixel.For each image block P, pass through block
Method of completing the square searches for similar image block in noise sequence of frames of video, and obtains m similar image block in each frame;For each
Its all row is in turn connected to form a column vector, then the column vector collection of all similar blocks is combined into by similar image block:{di}(i
=1,2 ..., mn).On this basis, observation block matrix is constituted with all similar block column vectorsWherein D's is each
Row are all a similar block column vector, Rh×wIt is the real number matrix set that h, columns are w to represent line number.
Image block matching method in the step one, is including but not limited to entered based on Euclidean distance between image block vector
The method of row image Block- matching;
Step 2:Construct Poisson-Gaussian mixed noise priori;Its concrete form is:
Wherein, F (A, E, Z) represents Poisson-Gaussian mixed noise priori;A is the block matrix that clear block column vector is constituted, A
Each be classified as the corresponding clear block column vector of foregoing similar block;E represents block-matching error block matrix, and each of E is classified as foregoing
The corresponding matching error column vector of similar block;Z represents impulsive noise matrix, and each of Z is classified as the corresponding pulse of foregoing similar block
Noise vector;D represents the observation block matrix that similar block column vector is constituted;Max (X) representing matrixs X greatest member, division arithmetic
It is that matrix is carried out by element with square root calculation;σ is mixes in Poisson, gaussian sum poisson noise, the standard of Gaussian noise part
Difference.And N all elements are all 1.
Step 3:The video denoising model that construction is set up under mixing Poisson, gaussian sum impulse noise environment;Concrete form
For:
Solve denoising block matrix A, the error block matrix E and impulsive noise matrix Z for causing target function value to reach minimum:
Wherein, A is the block matrix that clear block column vector is constituted, and E represents block-matching error block matrix, and Z represents impulsive noise
Matrix, F (A, E, Z) represents Poisson-Gaussian mixed noise priori, | | | |*, | | | |1With | | | |FRespectively matrix core model
Number, l1Norm and Frobenius norms, λ, μ and η are constant coefficient.After solving picture rich in detail block matrix A, A row to
Amount is remapped as s × s image blocks, obtains the corresponding picture rich in detail block of similar image block;
Step 4:Denoising model is solved using optimized algorithm, the picture rich in detail block after denoising is obtained;Solve picture rich in detail block
Optimized algorithm, include but are not limited to existing algorithm Accelerated Proximal Gradient (APG).
Step 5:Picture rich in detail merged block after denoising is generated into clear video image.Its specific method is:
For kth frame (k=1,2 ..., n) clear video image, if including pixel pkThe picture rich in detail set of blocks of (x, y)
For { b1,b2,…,bv}.Wherein, x, y are respectively line number and columns of the pixel in video image, pk(x, y) is (x, y) place
Pixel, v is includes pixel pkThe image block number of (x, y), { b1,b2,…,bvRepresent to include pixel pk(x's, y) is clear
Image block set.Then pixel pkThe gray value of (x, y) is:
Wherein, I (pk(x, y)) it is pixel pkThe gray value of (x, y),For bj(j=1,2 ..., v) in pixel
pk(x, y) corresponding gray value, v is to include pixel pkThe image block number of (x, y).
In concrete operations, using bus, coastguard, flower, missa and salesman video sequences [4] are carried out
Experimental evaluation.Poisson noise, Gaussian noise and obedience [0-255] equally distributed pulse is added on clear video frame images to make an uproar
Sound, then carries out denoising to noise frame of video, examines this method effect, and with the video denoising algorithm that is widely used at present:
VBM3D [2], VBM4D [3] and RPCA [4] are compared.Wherein, denoising after-vision effect (Gaussian noise relatively more as shown in Figure 2
Standard deviation is σ=20, and impulsive noise ratio is r=0.1), be respectively from left to right:Clear noiseless video image;Add pool
Noise video image after pine, gaussian sum impulsive noise;VBM3D algorithm denoising effects;VBM4D algorithm denoising effects;RPCA
Algorithm denoising effect;This method denoising effect.PSNR Indexes Comparisons after denoising are as shown in figure 3, wherein Gaussian noise standard deviation sigma is
20-40, impulsive noise ratio r are 10%-30%.By contrast as can be seen that several videos of this method and current main flow are gone
Method for de-noising, which is compared, can obtain more preferably mixed noise denoising effect.
One of ordinary skill in the art will appreciate that realizing all or part of flow in above-described embodiment method.In above
Hold a kind of described only preferred embodiment of the invention, the interest field of the present invention can not be limited with this certainly, because
This equivalent variations made according to the claims in the present invention, still belongs to the scope that the present invention is covered.