CN108376259A - In conjunction with the image denoising method of Bayes's Layered Learning and empty spectrum joint priori - Google Patents

In conjunction with the image denoising method of Bayes's Layered Learning and empty spectrum joint priori Download PDF

Info

Publication number
CN108376259A
CN108376259A CN201810064821.XA CN201810064821A CN108376259A CN 108376259 A CN108376259 A CN 108376259A CN 201810064821 A CN201810064821 A CN 201810064821A CN 108376259 A CN108376259 A CN 108376259A
Authority
CN
China
Prior art keywords
noise
bayes
image
model
priori
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.)
Pending
Application number
CN201810064821.XA
Other languages
Chinese (zh)
Inventor
刘帅
田智强
王昱童
彭思继
郑帅
李垚辰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian Jiaotong University
Original Assignee
Xian Jiaotong University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xian Jiaotong University filed Critical Xian Jiaotong University
Priority to CN201810064821.XA priority Critical patent/CN108376259A/en
Publication of CN108376259A publication Critical patent/CN108376259A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/29—Graphical models, e.g. Bayesian networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10032—Satellite or aerial image; Remote sensing
    • G06T2207/10036—Multispectral image; Hyperspectral image

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Processing (AREA)

Abstract

The invention discloses the high spectrum image noise-reduction methods of a kind of combination Bayes Layered Learning and empty spectrum joint priori, it is first depending on the empty spectrum correlation of high spectrum image and non local self-similarity, three-dimensional sliding block segmentation is carried out to it, and come non local is selected with block number to be observed according to several most like block numbers according to as collaboration block number evidence using the relative distance priori based on fusion feature；Secondly Bayes's low-rank decomposition model is built using layering priori, realize collaboration block number according to study and indicated.The model portrays the statistical property of collaboration block number evidence using low-rank decomposition, and expresses noise statistics by the Gaussian mixtures in conjunction with Di Li Cray processes；Then model is solved using variational Bayesian method, picture noise this purpose is effectively reduced to realize.The present invention not only solve the prior art for and meanwhile inhibit the deficiencies of a variety of this problem of noise in high spectrum image, and result is accurate, and strong analysis foundation is provided for its subsequent applications.

Description

In conjunction with the image denoising method of Bayes's Layered Learning and empty spectrum joint priori
Technical field
The invention belongs in information technology field, and in particular to a kind of combination Bayes Layered Learning and empty spectrum joint priori High spectrum image noise-reduction method.
Background technology
High spectrum image has very strong empty spectrum correlation, is based on this characteristic, block study is in high spectrum image analysis A large amount of application is obtained.But the position that all pieces are placed on equity by most of noise-reduction methods based on block study is learned It practises, has ignored the otherness of edge and partial structurtes between high-spectrum imaging different masses data, therefore cannot be to EO-1 hyperion Image carries out effective expression.Especially when image pixel in the block is seriously polluted by noise, the image is in the block effectively at this time Information is seldom, thus directly can not restore to obtain without data of making an uproar using the image block.On the other hand, existing research shows that it is high Spectrum picture is usually by the completely different noise pollution of a variety of statistical properties, such as dependent on the noise of signal, dependent on space The noise and mixed noise in domain between domain or spectrum.Existing majority Denoising Algorithm is analyzed either just for partial noise Noise with different statistical properties is separately modeled, the intrinsic noise characteristic of high spectrum image cannot be effectively characterized, drop It is showed above the adequacy made an uproar poor.
Invention content
The purpose of the present invention is to provide the high spectrum images of a kind of combination Bayes Layered Learning and empty spectrum joint priori On the one hand noise-reduction method, this method remain in image " totally " effective information, while can be to different noise statistics Effective expression, and then realize being sufficiently separated for image effective information and noise.On the other hand, this method utilizes variation Bayes Built statistical model is solved, optimal solution can be searched in the parameter space wirelessly tieed up, ensured the validity of this method And feasibility.
In order to achieve the above objectives, the technical solution adopted by the present invention is as follows：
In conjunction with the high spectrum image noise reduction of Bayes's Layered Learning and empty spectrum joint priori, include the following steps：
1) the collaboration block learning strategy using high spectrum image sky spectrum correlation and non local self-similarity is established；
2) Bayes's low-rank decomposition model is built using layering priori, to the system of effective information in high spectrum image and noise Meter characteristic is given full expression to；
3) deduction solution is carried out to correlated variables in model using variational Bayesian method.
The present invention, which further improves, to be, be directed to high spectrum image X, designed collaboration block learns plan in step 1) Slightly overlapping sliding block is carried out to X to divide, obtain center in sample x using the three-dimensional bits that size is d × d × λiBlock number according to N (xi)。 For each sample xi, N (xi) it can be regarded as that there is the set that the pixel of empty spectrum similitude is formed by one group.Define yi It indicates by block number according to N (xi) in data combination made of vector.Using the mean value of all elements in each image block as the figure As the fusion feature of block, the similitude between different masses data then is calculated according to the fusion feature, characteristics of mean should be based on Relative distance priori it is as follows：
Wherein,It is the d that element value is 12λ ties up row vector.It will be evident that SimilarIndex (yi,yj) value get over Greatly, then it represents that block number is according to N (xi) and N (xj) between it is more similar.According to SimilarIndex (yi,yj), it selects and data yiMost phase As (P-1) a non local block number evidence, and then realize sample xiCollaboration block YiStructure.In order to calculate simplicity, usually in pixel xiW × W neighborhoods in calculate and yiMost like P block number evidence；When W is sufficiently large, it is believed that be in whole picture high spectrum image Search and yiMost like data.
The present invention, which further improves, to be, in step 2), Bayes's low-rank decomposition model is built using layering priori, real Effective information and noise statistics gives full expression in existing high spectrum image.In order to express conveniently, Y is replaced using symbol Yi, Established low-rank decomposition model is expressed as：
Y=UVT+N
First item indicates low-rank decomposition item.Wherein, U=[u1,…,ui,…,uL] and V=[v1,…,vj,…,vP]TTo divide Dematrix.ui∈RMAnd vj∈RLIt is zero to obey mean value respectively, variance τuiAnd τvjGaussian Profile.In order to improve the steady of model Strong property and reduction are to parameter τuiAnd τvjSensibility, introduce gamma distribution and it constrained.The following institute of mathematic(al) representation Show：
τui~Γ (a0,b0), τvj~Γ (c0,d0)
Wherein, a0、b0、c0And d0For the hyper parameter of low-rank decomposition item.
Section 2 N=[n1,…,ni,…,nP] indicate noise item.In view of there are a variety of statistical properties in high spectrum image Completely different noise, and MOG models can effectively learn and it is arbitrary continuously distributed to characterize, thus using MOG models come Noise is indicated.It is defined as follows：
μk~N (μ0,Σ0)Σk~iWishart (e0,f0)
Wherein, μkAnd ΣkIt indicates the mean value and variance of k-th of Gaussian Profile, and is divided using Gaussian Profile and inverse Wishart Cloth comes to μkAnd ΣkLearnt and is constrained.Here, Gaussian Profile and inverse Wishart distributions meet conjugacy.μ0And Σ0 Expression parameter μkMean value and variance；e0And f0The degree of freedom and Scale Matrixes of inverse Wishart distributions are indicated respectively.In order to realize The number of Gaussian Profile is determined in a manner of data-driven and chooses suitable Gaussian Profile to express noise, is introduced Indicator variable zij∈{0,1}K, K → ∞, and ∑kzijk=1.zijIt obeys parameter and is the multinomial distribution of π, and utilize Dirichlet processes learn π.
In addition, for missing pixel, institute's extracting method dataIt is observed to substitute.Wherein Δ={ 0,1 }M×PIt is the sampling matrix that element value is 0 or 1.Therefore, Σfi=0 (f=1 ..., M) it indicates in gathered data yiWhen lose YiF-th of element；Σfi=1 (f=1 ..., M) then indicate data yiF-th of pixel effectively acquired.
The present invention, which further improves, to be, in step 3), using variational Bayesian method in the parameter space wirelessly tieed up In the correlated variables of constructed model in step 2) is solved, it is assumed that symbol Ψ={ ui,vi,μk,Σk,zi,τui,τvj, vtIndicate correlated variables in constructed Bayes's low-rank decomposition model；Symbol Θ={ a0,b0,c0,d0,μ0,Σ0,e0,f0,β} It indicates corresponding hyper parameter in model, in the case of given observation data Y and hyper parameter Θ, estimates hidden variable Ψ in model Posterior distrbutionp；In solution procedure, q (Ψ)=Π is distributed by variationfΨfTo approach the true Posterior distrbutionp p of hidden variable Ψ (ΨH,Θ).And then the more new formula of correlated variables in constructed low-rank decomposition model in step 2 is obtained, it is specific as follows：
(1) z is updatedijAnd vt：
vtPosterior distrbutionp be still beta distributions, and vt~beta (gt,ht), then have,
For variable zij,
Wherein,
Φ is enabled to indicate digamma functions,WithMeet,
(2) μ is updatedk：
μkPosterior distrbutionp be still Gaussian Profile, and meet,
(3) Σ is updatedk:
ΣkPosterior distrbutionp be still that inverse Wishart is distributed, and is met,
It can obtain,
(4) u is updatedi：
uiPosterior distrbutionp be still Gaussian Profile, and meet,
Wherein, mean valueAnd varianceFor,
(5) τ is updatedui：
τuiPosterior distrbutionp be still gamma distribution, and meet,
τui~Γ (a, bi)
Wherein, parameter a and biFor,
(6) v is updatedi：
viPosterior distrbutionp be still Gaussian Profile, and meet,
Wherein, mean valueAnd varianceFor,
(7) τ is updatedvj：
τvjPosterior distrbutionp be still gamma distribution, and meet,
τvj~Γ (c, dj)
Wherein, parameter c and djFor,
The present invention has the advantage that：
The present invention combines realization EO-1 hyperion under Bayesian frame, by the study of collaboration block and low-rank sparse decomposition model Image noise reduction this purpose.By cooperateing with block to learn come the empty spectrum correlation for excavating high spectrum image and non local self-similarity, Effective expression can be carried out to edge and the pixel to differ greatly in region；Learn and indicate collaboration using low-rank sparse decomposition The low-rank characteristic in empty spectral domain of block number evidence realizes the study to data structure information；Recycle variational Bayesian to mould Type is solved, and has the advantages that calculate simply and stability is high.
Further, research of the invention is related to statistical learning, collaboration indicates and multiple related necks such as management loading Domain has certain novelty, enriches existing image noise reduction theory, has important theory significance and practical value.
Description of the drawings
Fig. 1 for institute's extracting method of the present invention frame diagram；
Fig. 2 is algorithms of different to the Beads data sets the 30th that are polluted by Gaussian noise, sparse noise and missing pixel Wave band noise reduction result：
Fig. 2 (a) is reference picture；Fig. 2 (b) is noise image；Fig. 2 (c) is noise reduction result figure obtained by BM3D methods；Fig. 2 (d) it is noise reduction result figure obtained by ANLM3D methods；Fig. 2 (e) is noise reduction result figure obtained by BM4D methods；Fig. 2 (f) is LRMR methods Gained noise reduction result figure；Fig. 2 (g) is noise reduction result figure obtained by the method for the present invention；
Fig. 3 is noise reduction result of the algorithms of different to the 109th wave band of Urban data sets：
Fig. 3 (a) is true noisy image；Fig. 3 (b) is noise reduction result figure obtained by BM3D methods；Fig. 3 (c) is the side ANLM3D Noise reduction result figure obtained by method；Fig. 3 (d) is noise reduction result figure obtained by BM4D methods；Fig. 3 (e) is noise reduction result obtained by LRMR methods Figure；Fig. 3 (f) is noise reduction result figure obtained by the method for the present invention；
Fig. 4 is algorithms of different to the 90th wave band noise reduction knot of Pavia Centre data sets for being polluted by mixed noise Fruit：
Fig. 4 (a) is reference picture；Fig. 4 (b) is noise image；Fig. 4 (c) is noise reduction result figure obtained by BM3D methods；Fig. 4 (d) it is noise reduction result figure obtained by ANLM3D methods；Fig. 4 (e) is noise reduction result figure obtained by BM4D methods；Fig. 4 (f) is LRMR methods Gained noise reduction result figure；Fig. 4 (g) is noise reduction result figure obtained by the method for the present invention.
Specific implementation mode
It elaborates with reference to the accompanying drawings and examples to the present invention, but protection scope of the present invention is not limited to institute Embodiment.
The present invention makes improvement for existing high spectrum image noise-reduction method, it is proposed that a kind of combination Bayes layering Practise the high spectrum image noise-reduction method with empty spectrum joint priori.With reference to figure 1, the present invention excavates EO-1 hyperion by cooperateing with block study The empty spectrum correlation and non local self-similarity of image can carry out effective table to edge and the pixel to differ greatly in region It reaches；Learn and indicate the low-rank characteristic in empty spectral domain of collaboration block number evidence using low-rank decomposition, realizes to data structure information Study；And the noise characteristic of high spectrum image is simulated using the MOG noise modes in conjunction with Di Li Cray processes, it realizes Noise suppressed this purpose.Variational Bayesian is used to solve model, has calculating simply and stability is high Advantage.
The present invention is directed to the architectural characteristic of high spectrum image different masses data, and high spectrum image is divided into multiple overlappings Three-dimensional bits carry out image analysis.Different masses data represent the empty spectrum signature of difference of high spectrum image in high spectrum image, simultaneously Existing similitude between different images block, and have certain otherness.It is d × d first with size for high spectrum image X The three-dimensional bits of × λ carry out overlapping sliding block segmentation to X, obtain center in sample xiBlock number according to N (xi).Then for each sample This xi, N (xi) it can be regarded as that there is the set that the pixel of empty spectrum similitude is formed by one group.Define yiIt indicates by block number according to N (xi) in data combination made of vector.It is special using the mean value of all elements in each image block as the fusion of the image block Sign, finally calculates the similitude between different masses data according to the fusion feature, and it is first to be somebody's turn to do the relative distance based on characteristics of mean It tests as follows：
Wherein,It is the d that element value is 12λ ties up row vector.It will be evident that SimilarIndex (yi,yj) value get over Greatly, then it represents that block number is according to N (xi) and N (xj) between it is more similar.According to SimilarIndex (yi,yj), it selects and data yiMost phase As (P-1) a non local block number evidence, and then realize sample xiCollaboration block YiStructure.In order to calculate simplicity, usually in pixel xiW × W neighborhoods in calculate and yiMost like P block number evidence；When W is sufficiently large, it is believed that be in whole picture high spectrum image Search and yiMost like data.
On this basis, using layering priori structure Bayes's low-rank decomposition model to effective information in high spectrum image and Noise characteristic carries out statistical modeling.In order to express conveniently, Y is replaced using symbol Yi, established low-rank decomposition model is expressed as：
Y=UVT+N
First item indicates low-rank decomposition item.Wherein, U=[u1,…,ui,…,uL] and V=[v1,…,vj,…,vP]TTo divide Dematrix.ui∈RMAnd vj∈RLIt is zero to obey mean value respectively, variance τuiAnd τvjGaussian Profile.In order to improve the steady of model Strong property and reduction are to parameter τuiAnd τvjSensibility, introduce gamma distribution and it constrained.The following institute of mathematic(al) representation Show：
τui~Γ (a0,b0), τvj~Γ (c0,d0)
Wherein, a0、b0、c0And d0For the hyper parameter of low-rank decomposition item.
Section 2 N=[n1,…,ni,…,nP] indicate noise item.In view of there are a variety of statistical properties in high spectrum image Completely different noise, and MOG models can effectively learn and it is arbitrary continuously distributed to characterize, thus using MOG models come Noise is indicated.It is defined as follows：
μk~N (μ0,Σ0)Σk~iWishart (e0,f0)
Wherein, μkAnd ΣkIt indicates the mean value and variance of k-th of Gaussian Profile, and is divided using Gaussian Profile and inverse Wishart Cloth comes to μkAnd ΣkLearnt and is constrained.Here, Gaussian Profile and inverse Wishart distributions meet conjugacy.μ0And Σ0 Expression parameter μkMean value and variance；e0And f0The degree of freedom and Scale Matrixes of inverse Wishart distributions are indicated respectively.In order to realize The number of Gaussian Profile is determined in a manner of data-driven and chooses suitable Gaussian Profile to express noise, is introduced Indicator variable zij∈{0,1}K, K → ∞, and ∑kzijk=1.zijIt obeys parameter and is the multinomial distribution of π, and utilize Dirichlet processes learn π.
In addition, for missing pixel, present invention dataIt is observed to substitute.Wherein Δ={ 0,1 }M×P It is the sampling matrix that element value is 0 or 1.Therefore, Σfi=0 (f=1 ..., M) it indicates in gathered data yiWhen be lost yi F-th of element；Σfi=1 (f=1 ..., M) then indicate data yiF-th of pixel effectively acquired.
The correlated variables of constructed model is solved in the parameter space wirelessly tieed up using variational Bayesian method, Conventional letter Ψ={ ui,vi,μk,Σk,zi,τui,τvj,vtIndicate that the correlation in constructed Bayes's low-rank decomposition model becomes Amount；Symbol Θ={ a0,b0,c0,d0,μ0,Σ0,e0,f0, β } and indicate corresponding hyper parameter in model.In given observation data Y and In the case of hyper parameter Θ, the Posterior distrbutionp of hidden variable Ψ in model is estimated；In solution procedure, q (Ψ) is distributed by variation =ΠfΨfTo approach the true Posterior distrbutionp p (Ψ | H, Θ) of hidden variable Ψ.And then obtain phase in constructed low-rank decomposition model The more new formula of variable is closed, it is specific as follows：
(1) z is updatedijAnd vt：
vtPosterior distrbutionp be still beta distributions, and vt~beta (gt,ht), then have,
For variable zij,
Wherein,
Φ is enabled to indicate digamma functions,WithMeet,
(2) μ is updatedk：
μkPosterior distrbutionp be still Gaussian Profile, and meet,
(3) Σ is updatedk:
ΣkPosterior distrbutionp be still that inverse Wishart is distributed, and is met,
It can obtain,
(4) u is updatedi：
uiPosterior distrbutionp be still Gaussian Profile, and meet,
Wherein, mean valueAnd varianceFor,
(5) τ is updatedui：
τuiPosterior distrbutionp be still gamma distribution, and meet,
τui~Γ (a, bi)
Wherein, parameter a and biFor,
(6) v is updatedi：
viPosterior distrbutionp be still Gaussian Profile, and meet,
Wherein, mean valueAnd varianceFor,
(7) τ is updatedvj：
τvjPosterior distrbutionp be still gamma distribution, and meet,
τvj~Γ (c, dj)
Wherein, parameter c and djFor,
Emulation experiment and result
Fig. 2 shows algorithms of different to the Beads data sets the 30th that are polluted by Gaussian noise, sparse noise and missing pixel A wave band noise reduction result.Compared to noise image shown in Fig. 2 (b), the quality of Fig. 2 (c)-(g) has been significantly improved.According to According to Fig. 2 as can be seen that the method for the present invention effectively can realize denoising, and noise reduction result energy to the prodigious image of luminance difference Enough it is effectively maintained the information and object edge of homogenous region.BM3D methods press down image using Block- matching three-dimensional filtering It makes an uproar, noise reduction result has smoothed out some feature structures, is relatively obscured in visual effect.Recovery capabilities of the ANLM3D to detailed information All weaker, ANLM3D ratios BM3D is more excellent in visual effect, and the noise reduction result of BM4D is excessively smooth, is lost part edge Information.As shown in Fig. 2 (f), the noise reduction results of LRMR methods is substantially better than BM3D and ANLM3D, but the noise reduction knot of LRMR methods Fruit still has obvious sparse noise.Generally speaking, the method for the present invention can be good at removing the mixing in Beads data sets Noise, anti-acoustic capability have obviously superiority.
Noise reduction result of the algorithms of different to the 109th wave band of Urban data sets is set forth in Fig. 3.It can be with from Fig. 3 (a) Find out, the two by Banded improvement and mixed noise serious pollution.As can be seen that having obviously in Fig. 3 (b)-(d) Banded improvement, and picture structure and marginal information are relatively fuzzyyer, and this illustrates BM3D, ANLM3D and BM4D algorithm to by noise The noise reduction capability of serious pollution wave band is weaker.LRMR methods are more better to the recovery capability of image object and details, still Denoising result still has obvious Banded improvement and mixed noise.Find out from Fig. 3 (f), the method for the present invention is effectively inhibiting Banded improvement and mixed noise simultaneously, additionally it is possible to the detailed information such as edge and texture to image are effectively restored.
Fig. 4 gives algorithms of different to the 90th wave band noise reduction of Pavia Centre data sets for being polluted by mixed noise As a result.As can be seen that compared to noise image shown in Fig. 4 (b), BM3D, ANLM3D, BM4D, LRMR and the method for the present invention drop The picture quality for result of making an uproar, which is obtained for, to be obviously improved.From Fig. 4 (c) as can be seen that the obtained noise reduction result of BM3D methods its Structure is relatively fuzzyyer, and this method cannot effectively inhibit Banded improvement.These three methods of ANLM3D, BM4D and LRMR can only Suppressing portion divides Banded improvement.From Fig. 4 (g) as can be seen that the method for the present invention can effectively inhibit to Gaussian noise, sparse noise Inhibited with bad line, gained noise reduction is the result is that optimal.
In short, utilizing high spectrum image sky spectrum united information and noise characteristic these two aspects for existing noise-reduction method Deficiency, propose a kind of by Bayes's Layered Learning and high spectrum image noise-reduction method that empty spectrum joint priori is combined.Pass through Collaboration block learns to filter out with block number to be tested according to several most like block number evidences, is effectively composed to the sky of different masses data Correlation and non local self-similarity are portrayed, and promotion that can be fabulous is to seriously being polluted the noise reduction of block number evidence by noise Performance.Using low-rank decomposition to cooperateing with the low-rank characteristic in empty spectral domain of block number evidence to realize expression；By combining Di Li Crays The MOG noise modes of process to the noise characteristic of high spectrum image effectively portray.Variation Bayes calculates and is used to pair Model is solved, and has the advantages that calculate simple and stability.
Invention described above is only the preferred embodiment of the present invention, it should be pointed out that：For the common skill of the art For art personnel, without departing from the principle of the present invention, several improvements and modifications being expected can also be made, these Improvements and modifications also should be regarded as protection scope of the present invention.

Claims (4)

1. combining the image denoising method of Bayes's Layered Learning and empty spectrum joint priori, which is characterized in that include the following steps：
1) the collaboration block learning strategy using high spectrum image sky spectrum correlation and non local self-similarity is established；
2) Bayes's low-rank decomposition model is built using layering priori, it is special to the statistics of effective information in high spectrum image and noise Property is given full expression to；
3) deduction solution is carried out to correlated variables in model using variational Bayesian method.
2. the image denoising method of combination Bayes Layered Learning according to claim 1 and empty spectrum joint priori, special Sign is, the collaboration block learning strategy, and in step 1), for high spectrum image X, it is the three-dimensional of d × d × λ to utilize size Block carries out it overlapping sliding block segmentation, obtains center in sample xiBlock number according to N (xi), for each sample xi, N (xi) can There is the set that the pixel of empty spectrum similitude is formed by one group to regard as, define yiIt indicates by block number according to N (xi) in data group Vectorial made of conjunction, using the mean value of all elements in each image block as the fusion feature of the image block, then foundation should Fusion feature calculates the similitude between different masses data, should relative distance priori based on characteristics of mean it is as follows：
Wherein,It is the d that element value is 12λ ties up row vector, it will be apparent that, SimilarIndex (yi,yj) value it is bigger, then Indicate block number according to N (xi) and N (xj) between it is more similar, according to SimilarIndex (yi,yj), it selects and data yiMost like (P-1) a non local block number evidence, and then realize sample xiCollaboration block YiStructure, in order to calculate simplicity, usually in pixel xi's Calculating and y in W × W neighborhoodsiMost like P block number evidence；When W is sufficiently large, it is believed that searched in whole picture high spectrum image Rope and yiMost like data.
3. the image denoising method of combination Bayes Layered Learning according to claim 1 and empty spectrum joint priori, special Sign is, using Bayes's low-rank decomposition model constructed by layering priori, to effective information in high spectrum image and noise Statistical property is given full expression to, and after step 1), Y is replaced using symbol Yi, established low-rank decomposition model is expressed as：
Y=UVT+N
First item indicates low-rank decomposition item, wherein U=[u1,…,ui,…,uL] and V=[v1,…,vj,…,vP]TTo decompose square Battle array, ui∈RMAnd vj∈RLIt is zero to obey mean value respectively, variance τuiAnd τvjGaussian Profile, introduce gamma distribution comes to τuiWith τvjIt is constrained, to improve model robustness and reduce to parameter τuiAnd τvjSensibility, mathematic(al) representation is as follows：
τui~Γ (a0,b0), τvj~Γ (c0,d0)
Wherein, a0、b0、c0And d0For the hyper parameter of low-rank decomposition item；
Section 2 N=[n1,…,ni,…,nP] indicate noise item, it is distributed using MOG model learnings and characterization arbitrary continuation, in turn The expression for realizing a variety of completely different noises of statistical property present in high spectrum image, is defined as follows：
μk~N (μ0,Σ0) Σk~iWishart (e0,f0)
Wherein, μkAnd ΣkIndicate the mean value and variance of k-th of Gaussian Profile, and using Gaussian Profile and inverse Wishart be distributed come To μkAnd ΣkLearnt and constrained, here, Gaussian Profile and inverse Wishart distributions meet conjugacy, μ0And Σ0It indicates Parameter μkMean value and variance；e0And f0The degree of freedom and Scale Matrixes for indicating inverse Wishart distributions respectively, introduce instruction and become Measure zij∈{0,1}K, K → ∞, and ∑kzijk=1, zijThe multinomial distribution that parameter is π is obeyed, and utilizes Dirichlet processes To learn to π；
In addition, for missing pixel, institute's extracting method dataIt is observed to substitute, wherein Δ={ 0,1 }M×PIt is The sampling matrix that element value is 0 or 1, therefore, Σfi=0 (f=1 ..., M) it indicates in gathered data yiWhen be lost yiThe F element；Σfi=1 (f=1 ..., M) then indicate data yiF-th of pixel effectively acquired.
4. combining the image denoising method of Bayes's Layered Learning and empty spectrum joint priori, feature according to claim 1 It is, the method for solving based on variation Bayes of the step 3), it is assumed that symbol Ψ={ ui,vi,μk,Σk,zi,τui,τvj, vtIndicate Bayes's low-rank decomposition model correlated variables；Symbol Θ={ a0,b0,c0,d0,μ0,Σ0,e0,f0, β } and indicate model In corresponding hyper parameter estimate the posteriority point of hidden variable Ψ in model in the case of given observation data Y and hyper parameter Θ Cloth；In solution procedure, q (Ψ)=∏ is distributed by variationfΨfCome approach hidden variable Ψ true Posterior distrbutionp p (Ψ | H, Θ), and then the more new formula of correlated variables in constructed low-rank decomposition model in claim 2 is obtained, it is specific as follows：
(1) z is updatedijAnd vt：
vtPosterior distrbutionp be still beta distributions, and vt~beta (gt,ht), then have,
For variable zij,
Wherein,
Φ is enabled to indicate digamma functions,WithMeet,
(2) μ is updatedk：
μkPosterior distrbutionp be still Gaussian Profile, and meet,
(3) Σ is updatedk:
ΣkPosterior distrbutionp be still that inverse Wishart is distributed, and is met,
It can obtain,
(4) u is updatedi：
uiPosterior distrbutionp be still Gaussian Profile, and meet,
Wherein, mean valueAnd varianceFor,
(5) τ is updatedui：
τuiPosterior distrbutionp be still gamma distribution, and meet, τui~Γ (a, bi)
Wherein, parameter a and biFor,
(6) v is updatedi：
viPosterior distrbutionp be still Gaussian Profile, and meet,
Wherein, mean valueAnd varianceFor,
(7) τ is updatedvj：
τvjPosterior distrbutionp be still gamma distribution, and meet, τvj~Γ (c, dj)
Wherein, parameter c and djFor,
CN201810064821.XA 2018-01-23 2018-01-23 In conjunction with the image denoising method of Bayes's Layered Learning and empty spectrum joint priori Pending CN108376259A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810064821.XA CN108376259A (en) 2018-01-23 2018-01-23 In conjunction with the image denoising method of Bayes's Layered Learning and empty spectrum joint priori

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810064821.XA CN108376259A (en) 2018-01-23 2018-01-23 In conjunction with the image denoising method of Bayes's Layered Learning and empty spectrum joint priori

Publications (1)

Publication Number Publication Date
CN108376259A true CN108376259A (en) 2018-08-07

Family

ID=63016572

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810064821.XA Pending CN108376259A (en) 2018-01-23 2018-01-23 In conjunction with the image denoising method of Bayes's Layered Learning and empty spectrum joint priori

Country Status (1)

Country Link
CN (1) CN108376259A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110310234A (en) * 2019-05-17 2019-10-08 浙江工业大学 The facial image denoising method and device with Shannon fully differential are approached based on low-rank
CN111260576A (en) * 2020-01-14 2020-06-09 哈尔滨工业大学 Hyperspectral unmixing algorithm based on de-noising three-dimensional convolution self-coding network
CN112767271A (en) * 2021-01-19 2021-05-07 西安交通大学 Hyperspectral image depth noise reduction method based on three-dimensional variation network
CN113409307A (en) * 2021-07-15 2021-09-17 中南大学 Image denoising method, device and medium based on heterogeneous noise characteristics

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104268397A (en) * 2014-09-24 2015-01-07 杭州电子科技大学 Underwater acoustic target identifying method based on ship-radiated noise optimization features
CN106447630A (en) * 2016-09-18 2017-02-22 清华大学 High-spectral image sharpening method based on probability matrix decomposition
CN107436421A (en) * 2017-07-24 2017-12-05 哈尔滨工程大学 Mixed signal DOA estimation method under a kind of management loading framework

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104268397A (en) * 2014-09-24 2015-01-07 杭州电子科技大学 Underwater acoustic target identifying method based on ship-radiated noise optimization features
CN106447630A (en) * 2016-09-18 2017-02-22 清华大学 High-spectral image sharpening method based on probability matrix decomposition
CN107436421A (en) * 2017-07-24 2017-12-05 哈尔滨工程大学 Mixed signal DOA estimation method under a kind of management loading framework

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
刘帅等: "结合分层字典学习和空谱信息的多光谱图像去噪", 《西安电子科技大学学报（自然科学版）》 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110310234A (en) * 2019-05-17 2019-10-08 浙江工业大学 The facial image denoising method and device with Shannon fully differential are approached based on low-rank
CN111260576A (en) * 2020-01-14 2020-06-09 哈尔滨工业大学 Hyperspectral unmixing algorithm based on de-noising three-dimensional convolution self-coding network
CN111260576B (en) * 2020-01-14 2022-07-05 哈尔滨工业大学 Hyperspectral unmixing algorithm based on de-noising three-dimensional convolution self-coding network
CN112767271A (en) * 2021-01-19 2021-05-07 西安交通大学 Hyperspectral image depth noise reduction method based on three-dimensional variation network
CN112767271B (en) * 2021-01-19 2023-09-22 西安交通大学 Method for deep noise reduction of hyperspectral image based on three-dimensional variational network
CN113409307A (en) * 2021-07-15 2021-09-17 中南大学 Image denoising method, device and medium based on heterogeneous noise characteristics

Similar Documents

Publication Publication Date Title
Zhang et al. Edge strength similarity for image quality assessment
CN108376259A (en) In conjunction with the image denoising method of Bayes's Layered Learning and empty spectrum joint priori
Xu et al. Image inpainting by patch propagation using patch sparsity
CN110796625B (en) Image compressed sensing reconstruction method based on group sparse representation and weighted total variation
CN106462771A (en) 3D image significance detection method
CN103150713A (en) Image super-resolution method of utilizing image block classification sparse representation and self-adaptive aggregation
Wang et al. Centralized collaborative sparse unmixing for hyperspectral images
CN105046664A (en) Image denoising method based on self-adaptive EPLL algorithm
CN106251320A (en) Remote sensing image fusion method based on joint sparse Yu structure dictionary
US9183671B2 (en) Method for accelerating Monte Carlo renders
CN103208097A (en) Principal component analysis collaborative filtering method for image multi-direction morphological structure grouping
CN105160647A (en) Panchromatic multi-spectral image fusion method
CN103971354A (en) Method for reconstructing low-resolution infrared image into high-resolution infrared image
CN104657951A (en) Multiplicative noise removal method for image
CN104252704A (en) Total generalized variation-based infrared image multi-sensor super-resolution reconstruction method
Fan et al. Multiscale cross-connected dehazing network with scene depth fusion
Wang et al. Single image haze removal via attention-based transmission estimation and classification fusion network
Huang et al. SIDNet: a single image dedusting network with color cast correction
An et al. Image super-resolution reconstruction algorithm based on significant network connection-collaborative migration structure
CN107085826A (en) Based on the non local single image super resolution ratio reconstruction method for returning priori of weighted overlap-add
CN111539434B (en) Infrared weak and small target detection method based on similarity
Yang et al. Detail-aware near infrared and visible fusion with multi-order hyper-Laplacian priors
Yuan et al. Single-image rain removal using deep residual network
CN108776958B (en) Mix the image quality evaluating method and device of degraded image
Guan et al. DiffWater: Underwater Image Enhancement Based on Conditional Denoising Diffusion Probabilistic Model

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
WD01 Invention patent application deemed withdrawn after publication
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20180807