CN106023109A - Region similar example learning-based sparse denoising method - Google Patents

Abstract

The invention discloses a region similar example learning-based sparse denoising method. The method comprises the following steps: carrying out SIFT-MSER feature cluster-based similar region search in a picture library by using an image which contains noise, and calculating a corresponding affine transformation matrix by utilizing a matched SIFT feature point coordinate after obtaining similar regions; converting the matched similar regions to a direction and a size same as that of the image which contains noise; taking the similar regions after the affine transformation as examples of dictionary learning so as to improve the relevance between a dictionary and the image which contains noise; and finally carrying out high-frequency compensation to improve the texture information of the denoised image. According to the method disclosed in the invention, good denoising effect is obtained, the over-complete learning examples are changed, the learning examples are obtained through selecting similar regions, and the atomic structures of over-complete dictionaries are influenced, so that the atoms in the dictionaries can better express the effective parts of the images and neglect the noise parts; and through high-frequency compensation, the information of the similar images is utilized and the signal to noise ratio of the denoised images is enhanced.

Description

A kind of sparse denoising method based on the study of the similar sample in region

Technical field

The invention belongs to image and go dry processing technology field, particularly relate to a kind of based on the study of the similar sample in region sparse Denoising method.

Background technology

Digital picture in reality is subjected to imaging device in digitized and transmitting procedure and disturbs with external environmental noise Deng impact, the most noisy image or noise image.The process of noise in digital picture that reduces is referred to as image denoising.Removal image is made an uproar The method of sound includes: fold mean filter, uses the mean filter of neighborhood averaging to be highly suitable for removing by scanning Grain noise in the image obtained.Field averaging method effectively inhibits noise, as well as averagely causing fuzzy Phenomenon, fog-level is directly proportional to field radius.The smoothness that geometric mean wave filter is reached can filter with arithmetic equal value Device is compared, but can lose less pictorial detail in filtering.Harmonic wave mean filter is more preferable to " salt " noise effects, but It is not to be suitable for " Fructus Piperis " noise.It is good at processing other noises as Gaussian noise.Inverse harmonic wave mean filter is suitableeer Together in processing impulsive noise, but it has individual shortcoming, it is simply that have to know that noise is dark noise or bright noise, in order to select Suitably filter order numerical symbol, if the symbol of exponent number has selected wrong may cause catastrophic consequence.Fold self adaptation Wiener filter, it can adjust the output of wave filter according to the local variance of image, and local variance is the biggest, wave filter smooth Act on the strongest.Its final goal be make recovery image f^ (x, y) with original image f (x, mean square error e2=E y) [(f (x, Y)-f^ (x, y) 2] minimum.The filter effect of the method is better than mean filter effect, to retain the edge of image and other HFS is very useful, but amount of calculation is bigger.The Wiener filter imaging filtering best results to having white noise.In folding Value filter, it is a kind of conventional Nonlinear Smoothing Filter, and its ultimate principle is in digital picture or Serial No. one The value of point is to allow the difference of surrounding pixel gray value compare by its major function that replaces of the intermediate value of each point value in a field of this point Big pixel changes to take the value close with the pixel value of surrounding, such that it is able to eliminate isolated noise spot, so medium filtering for The salt-pepper noise filtering image is highly effective.Median filter can be accomplished not only to remove noise but also can protect the edge of image, from And obtain relatively satisfactory recovery effect, and, during actual operation, need not the statistical property of image, this also brings much Convenient, but many to some details, that image that particularly point, line, pinnacle details are more should not use medium filtering method.Fold Morphology scratch filter, combines unlatching and Guan Bi and can be used to filter noise, first to there being noisy image to open Operation, optional structure salt matrices is bigger than the size of noise, thus the result opened is by the noise remove in background.Finally It is that the image obtaining back carries out closed procedure, the noise on image is removed.Feature according to the method it is recognised that The image type that the method is suitable for is that the object size in image is the biggest, and does not has tiny details, to such The effect of scene image partition can be relatively good.Folding Wavelet Denoising Method, this method remains the wavelet coefficient that major part comprises signal, because of This can preferably keep pictorial detail.Wavelet analysis carries out image denoising mainly 3 steps: picture intelligence is carried out by (1) Wavelet decomposition.(2) high frequency coefficient after hierachical decomposition is carried out threshold value quantizing.(3) 2-d wavelet reconstructed image is utilized to believe Number.

Prior art does not utilize the prior information making image object, and these algorithms are all based on noise and meet fixing dividing Cloth model, such as Gauss distribution, and real system imaging has bigger gap, and algorithm universality is poor, and these algorithms are at preferable noise mode There is under type preferable denoising effect, but for the real image by multiple actual noise interactive interference, then under effect is notable Fall.

Summary of the invention

It is an object of the invention to provide a kind of sparse denoising method based on the study of the similar sample in region, it is intended to solve In the case of known portions imageable target object part prior information, target image high accuracy goes to make problem.Based on Aerospace Satellite Target has similar material, architectural feature, and such as solar energy sailboard, sensor, satellite main body etc., the present invention utilizes the most known Three-dimensional artificial satellite model generation training dataset, found the local of denoising image object by SIFT-MSER object matching Region sample, and complete high precision image go to make with target area sample training dictionary, solve existing method dictionary and image The problem not having dependency.

The present invention is achieved in that a kind of sparse denoising method based on the study of the similar sample in region, described based on district The sparse denoising method of territory similar sample study:

Image first by Noise carries out similar area based on SIFT-MSER feature bunch retrieval in image library, The SIFT feature point coordinates of coupling is utilized to calculate corresponding affine transformation matrices after obtaining similar area；

The similar area conversion that will match to is to the direction identical with noise image and yardstick；

Finally use the similar area after affine transformation as the sample of dictionary learning, improve the phase of dictionary and noise image Guan Xing；Dictionary carries out high frequency compensation after going to make.

Further, the acquisition of similar area specifically includes:

The MSER feature extraction of image, after setting up sample image data set Ω, is obtained by MSER algorithm and SIFT algorithm Fetch data and collect maximum stable extremal region and the scale invariant feature descriptor of intra-zone of every piece image in Ω；Each width Image I presses specific structure after calculating MSER region and SIFT spy's feature and stores in file；Gray level image is carried out two Value obtains the binaryzation sequence of a width figure；

Build the SIFT-MSER characteristic area bunch of image, first extract MSER region, in MSER region, then extract SIFT Feature, M is the set in all MSER regions of image I, and S is the set of all SIFT feature of image I, all characteristic areas bunch It is denoted as:

Λ=< { m₁,S₁},{m₂,S₂},…,{m_n,S_n}>；

M=< { m₁},{m₂},…,{m_n}>；

S=< { f₁},{f₂},…,{f_n}>；

Wherein, f_j={ v_j,l_j,s_j,o_jRepresent a SIFT feature descriptor, v_jIt is that this SIFT key point is at l_j= (x_j,y_j) 128 dimensional vector descriptors of position, s_jIt is the value of yardstick, o_jIt it is the value in direction.S_j∈ S is a son of S Collection, represents to fall the jth maximum stable extremal region m at image I_jInternal all SIFT feature describe operator；s_j={ b_j, l_j,a_j,b_jRepresent the region of MSER, b_jIt is the collection of pixels of this characteristic area, l_jIt is positional information, a_j,b_jTable respectively Show the major and minor axis of ellipse；

Image local similar area based on SIFT-MSER feature bunch obtains, and closes setting up SIFT-MSER feature gathering After, in syndrome setIn find out and noise image I_sIth feature bunch Λ_iThe feature bunch that ∈ Λ is similar.

Further, Q is used in the acquisition of described similar area₁, Q₂, Q_i-1, Q_iRepresent this series of mutually nested extreme values Regional sequence:

$q\left(i\right)=\frac{\left|Q_{i+\mathrm{\&Delta;}}\right|-\left|Q_{i-\mathrm{\&Delta;}}\right|}{\left|Q_i\right|};$

Wherein i represents different threshold values, and Q is the set of a pixel, and what absolute value represented is this cardinality of a set, will It is as the area of this extremal region.

Further, the described definition according to Euclidean distance, the distance defined between two SIFT-MSER features bunch is:

$Dist({\mathrm{\&Lambda;}}_a,{\mathrm{\&Lambda;}}_b)=\sqrt{{\left(\frac{mean\left(M_a\right)-mean\left(M_b\right)}{255}\right)}^2+{\left(\frac{\left|S_a\right|+\left|S_b\right|}{2\&times;\left|match(S_a,S_b)\right|}\right)}^2};$

What absolute value represented is this cardinality of a set.mean(M_a) represent, match (S_a,S_b) represent S_aAt S_bIn have coupling The set of element, obtains S_a,S_bThe SIFT feature descriptor of middle coupling.Each MSER region for noise image obtains C, takes the region of C=10 coupling, then arrives the MSER regional ensemble Ψ of coupling.

Further, the process of study sample uses RANSAC algorithm, and concrete grammar is as follows:

Input: by observation data D of noise severe contamination, parameterized mathematical model M, the unknown ginseng of parameterized model Number m₁, m₂..., m_n；

Process: perform (1), (2), (3), (4) (5)

(1) selecting one group of data, at random and be assumed to be intra-office point, other data are assumed to be point not in the know；

(2) all parameters m of parameterized model, are calculated by these intra-office points₁, m₂..., m_n；

(3), by parameter m obtained₁, m₂..., m_nGo whether the point not in the know testing other meets parameterized model M, if Certain point meets parameterized model M, and this point is moved into intra-office point by that；

(4), use the similar approach such as least square, reappraise point in the owning administration of model, if this model ratio is existing Model have less error rate, then replace "current" model with model；

(5) if "current" model has sufficiently small error rate, or maximum iteration time has been reached, then algorithm terminates, Otherwise, (1), (2), (3), (4) are performed.

Further,That a DSIFT coupling is right, then the image block after high frequency compensation:

$I_{bc}=I_{b1}+\mathrm{\&gamma;}*(I_{b1}-I_{bx}^{*});$

Wherein I_b1Be original picture block to be compensated,For the noise-free picture block matched, γ is weight, takes here Empirical value 0.25.

The present invention provide based on the similar sample in region study sparse denoising method, not only obtain than BLS-GSM, Y-PSNR that NLM is high and in visual effect, also surmounted other two kinds of methods.The Denoising Algorithm frame that the present invention proposes Why frame can obtain good denoising effect, and key reason is to change the study sample of complete dictionary, by choosing Select similar area and obtain study sample, affect the atomic structure of complete dictionary so that the atom in dictionary can more preferable table The live part of diagram picture, ignores noise section.By high frequency compensation further with the information of similar image, further Improve the signal to noise ratio of denoising image.

Accompanying drawing explanation

Fig. 1 is the sparse denoising method flow chart based on the study of the similar sample in region that the embodiment of the present invention provides.

Fig. 2 is the affine transformation relationship schematic diagram that the RANSAC that the embodiment of the present invention provides estimates.

Fig. 3 is the round schematic diagram that the RANSAC that the embodiment of the present invention provides estimates.

Fig. 4 is first group of artificial satellite imaging experiments result PSNR schematic diagram that the embodiment of the present invention provides.

Fig. 5 is second group of artificial satellite imaging experiments result PSNR schematic diagram that the embodiment of the present invention provides.

Detailed description of the invention

In order to make the purpose of the present invention, technical scheme and advantage clearer, below in conjunction with embodiment, to the present invention It is further elaborated.Should be appreciated that specific embodiment described herein, only in order to explain the present invention, is not used to Limit the present invention.

The present invention is not in the case of having threedimensional model, it is possible to use three-dimensional artificial satellite structure dictionary instruction known to other Practice collection.The training of complete dictionary was carried out according to the regional area similarity of satellite shape.In reality, a lot of images are in some district Territory is similar, and combines SIFT feature and carry out the acquisition of similar sample with maximum stable extremal region (MSER).

Below in conjunction with the accompanying drawings the application principle of the present invention is further described.

As it is shown in figure 1, the sparse denoising method based on the study of the similar sample in region of the embodiment of the present invention includes following step Rapid:

S101: the image first by Noise carries out similar area based on SIFT-MSER feature bunch in picture library Retrieval, utilizes the SIFT feature point coordinates of coupling to calculate corresponding affine transformation matrices after obtaining similar area；

S102: the similar area conversion that will match to is to the direction identical with noise image and yardstick；

S103: finally use the similar area after radiation conversion as the sample of dictionary learning, improve dictionary and noise pattern The dependency of picture；Carry out high frequency compensation.

Below in conjunction with specific embodiment, the application principle of the present invention is further described.

The acquisition of 1 similar area

In order to obtain the description of similar area, the feature of the SIFT obtaining a region that SIFT and MSER is combined Bunch, and use this feature bunch to weigh the similarity in two MSER regions.The three-dimensional artificial satellite model using STK produces not The sample image data set Ω of multi-pose multi-angle with the artificial satellite of type.

The MSER feature extraction of 1.1 images

After setting up sample image data set Ω, obtain each width figure in data set Ω by MSER algorithm and SIFT algorithm The maximum stable extremal region of picture and the scale invariant feature descriptor of intra-zone.In order to improve algorithm operation efficiency avoid right Piece image repeats to extract maximum stable extremal region and scale invariant feature descriptor, and every piece image I calculates MSER region Store in file with pressing specific structure after SIFT spy's feature.

MSER (maximallystableextremeregion) i.e. maximum stable extremal region, has Affinely invariant region There is fine performance, be commonly used to as image retrieval, coupling service.Same object is often had and repeatedly claps from many different angles The picture taken the photograph, MSER can be used to judge whether to comprise in different photos same object.

Watershed concept in the similar landform of MSER use carries out binaryzation and obtains the binaryzation of a width figure gray level image Sequence.The three dimensional display of image, takes different value in Z-direction and image binaryzation can be obtained different bianry images, and Fig. 2 crosses out The binaryzation sequence of piece image.

Region: be exactly the part in image, any two point in the zone can arrive along adjacent pixel There is a path connected between another point, i.e. any two point, and the institute on path is the most also in region.

Extremal region: the threshold value that the condition of composition extremal region is as setting is continually changing the new vegetarian refreshments energy no longer having Enough it is incorporated to current region and makes current region expand.

To a secondary gray-scale map, choose a series of gray threshold, image is carried out binaryzation, such as, take T=0,1, 2 ..., 255, when threshold value is t, obtain binary image I_t.As in figure 2 it is shown, by the image sequence of all binaryzations by threshold value Sequentially show, it is found that along with engendering stain or white point on the modified-image of threshold value, stain gradually merges one Rise and form a little region, become connected component.The most respectively using black and white region as target area, these figures In, all of connected component is all extremal region.Even if in the picture, an isolated stain is also likely to be extremal region.Pole The determination in value region and gray threshold are closely related.For more region being detected, need original image is inverted, weight Recheck survey process, so make the black objects in image and white object can be detected as extremal region.

Maximum stable extremal region: in the extremal region found, some extremal region is along with the gray threshold arranged Increase and a little " growing up ".Such one group of extremal region is ascending is mutually nested relation, and these become nest relation Extremal region constitute Component Tree.Use Q₁, Q₂, Q_i-1, Q_iRepresent this series of mutually nested extremal regions Sequence.

$q\left(i\right)=\frac{\left|Q_{i+\mathrm{\&Delta;}}\right|-\left|Q_{i-\mathrm{\&Delta;}}\right|}{\left|Q_i\right|}---\left(1\right)$

Wherein i represents different threshold values, and Q is the set of a pixel, and what absolute value represented is this cardinality of a set (unit The number of element), it can be regarded the area of this extremal region.The Q when q (i) obtains minima_iJust it is properly termed as MSER. From the explanation above it can be seen that this maximum stable extremal region is when gray threshold changes, region area changes Minimum extremal region is exactly MSER.So first to find all of extremal region, then determine final by formula 1 MSER。

When image generation affine transformation, the gray value of pixel is impregnable, does not thus interfere with extreme value The acquisition in region.The viewpoint of geometry classification is thought, the collimation of two straight lines, the ratio of parallel segment, the ratio of closed figure area Deng, it is all to keep constant under affine transformation, so maximum stable extremal region is impregnable for q (i), is entering After row affine transformation, q (i) still can get minima at maximum stable extremal region, so this extraction characteristic area Method there is affine-invariant features.

When after the MSER region detection completing image for these irregular areas of Unify legislation, so these must be extracted Region corresponding shape and structure feature, the key message of a region shape structure can be attributed to position, size, direction, will Image-region fits to these key messages of expression that ellipse can be appropriate.The center of gravity of oval central representation image-region, also Being exactly the positional information in region, oval major and minor axis may determine that the size and Orientation of image-region.

The 1.2 SIFT-MSER characteristic areas bunch building image

In the algorithm, first extract MSER region, in MSER region, then extract SIFT feature, and by simple for two features Bind together formation SIFT-MSER characteristic area bunch, use this characteristic area bunch as the feature description in MSER region Collection, it is assumed that M is the set in all MSER regions of image I, and S is the set of all SIFT feature of image I, all characteristic areas Bunch can simply be denoted as Λ:.

Λ=< { m₁,S₁},{m₂,S₂},…,{m_n,S_n}> (2)

M=< { m₁},{m₂},…,{m_n}> (3)

S=< { f₁},{f₂},…,{f_n}> (4)

Wherein, f_j={ v_j,l_j,s_j,o_jRepresent a SIFT feature descriptor, v_jIt is that this SIFT key point is at l_j= (x_j,y_j) 128 dimensional vector descriptors of position, s_jIt is the value of yardstick, o_jIt it is the value in direction.S_j∈ S is a son of S Collection, represents to fall the jth maximum stable extremal region m at image I_jInternal all SIFT feature describe operator.s_j={ b_j, l_j,a_j,b_jRepresent the region of MSER, b_jIt is the collection of pixels of this characteristic area, l_jIt is positional information, a_j,b_jTable respectively Show the major and minor axis of ellipse.

It appeared that different characteristic areas bunch has mutually nested relation, the relation of the most this nesting adds weight Folded region ratio in dictionary training set so that the dictionary that study obtains preferably can carry out Its Sparse Decomposition to key area.

Generally when application, a lot of SIFT feature points can be hovered on the border of MSER, in order to avoid subtracting of SIFT feature point Few, artificial is multiplied by a factor 1.2 by the major and minor axis that MSER is oval, increases oval area, the most also can abandon those ellipse The MSER region that the area of a circle is the least, falls at MSER region r without any one SIFT key point_iIn, then this region The most to be discarded.

Multiple SIFT feature can be comprised by each SIFT-MSER characteristic area bunch, and MSER region exists overlap Phenomenon, so SIFT-MSER characteristic area bunch is more flexible, can allow less local matching.One artificial can be defended Star chart picture simply divides, and this point is somewhat like artificial satellite to be disassembled into less spare and accessory parts according to composition difference.

1.3 image local similar areas based on SIFT-MSER feature bunch obtain

After setting up the conjunction of SIFT-MSER feature gathering, next it is to find out in syndrome set and noise image I_s Ith feature bunch Λ_iThe feature bunch that ∈ Λ is similar.The method that algorithm uses Euclidean distance nearest in the present invention saves finds phase As characteristic block.

According to the definition of Euclidean distance, the distance defined between two SIFT-MSER features bunch is

$Dist({\mathrm{\&Lambda;}}_a,{\mathrm{\&Lambda;}}_b)=\sqrt{{\left(\frac{mean\left(M_a\right)-mean\left(M_b\right)}{255}\right)}^2+{\left(\frac{\left|S_a\right|+\left|S_b\right|}{2\&times;\left|match(S_a,S_b)\right|}\right)}^2}---\left(5\right)$

What absolute value represented is this cardinality of a set (number of element).mean(M_a) represent, match (S_a,S_b) represent S_a At S_bIn have coupling element set, obtain S_a,S_bThe SIFT feature descriptor of middle coupling.For each of noise image MSER region all obtains the region of C (taking C=10 in experiment) individual coupling, then arrives the MSER regional ensemble Ψ of coupling.

The process of 2 study samples

After obtaining similar area, next step calculates the homography matrix of matching area and noise image, uses in this step RANSAC algorithm calculates Λ_a,Λ_bBetween mapping relations.Make with the positional information of the SIFT descriptor of coupling between two MSER For calculating sample point.Obtaining Λ_a,Λ_bBetween mapping relations after, data set Ψ is carried out inverse transformation.Will match to MESR region is adjusted to the yardstick as noise MSER region, and the data after adjusting are stored in data set Ψ^*In.Data set Ψ^*In image-region be adjusted after and noise image on yardstick very close to, use Ψ^*In each image-region carry out 8 The sliding window of × 8 takes block operations and obtains the sample of dictionary learning.Obtained complete dictionary D by KSVD Algorithm Learning, then used OMP algorithm carries out Its Sparse Decomposition to noise image, reconstruct reduction obtains Pre-denoised image I_n ^*。

RANSAC (RANdomSAmple Consensus), stochastic sampling unification algorism.This algorithm is the earliest by Fischler Proposed in 1981 with Bolles, be widely used in later in computer vision, be used for estimating the basis matrix of stereo camera. To in world's modeling process, the phenomenon of various existence is abstracted into mathematical model.There are some parameters in each mathematical model, By different parameters, obtain different examples, deduce.For reality obtains a heap data by observation, how to be This heap data looks for a suitable model, then determines suitable model parameter, is a critically important problem, and RANSAC algorithm is just It it is one of them common derivation algorithm.

What in modeling the world, generation observed that data can be artificial is divided into valid data and invalid data two kinds.Due to Data acquisition is often affected by uncertain factor and makes data produce deviation, if the deviation produced is several too greatly According to cannot objectively reflect abstract mathematical model, these data are exactly invalid data.If invalid data is whole number Fraction according to collection, then the parameter of mathematical model can be solved by method of least square or similar method.If it is invalid The quantity of data has exceeded valid data, then now re-uses method of least square and solves and just become no longer valid, but RANSAC algorithm is the most effective.In RANSAC algorithm, valid data can be referred to as again intra-office point, and invalid data is referred to as point not in the know.

RANSAC arthmetic statement is as follows:

Input: by observation data D of noise severe contamination, parameterized mathematical model M, the unknown ginseng of parameterized model Number m₁, m₂..., m_n。

Process: perform (1), (2), (3), (4) (5)

(1) selecting one group of data, at random and be assumed to be intra-office point, other data are assumed to be point not in the know；

(2) all parameters m of parameterized model, are calculated by these intra-office points₁, m₂..., m_n；

(3), by parameter m obtained₁, m₂..., m_nGo whether the point not in the know testing other meets parameterized model M, if Certain point meets parameterized model M, and this point is moved into intra-office point by that.

(4), use the similar approach such as least square, reappraise point in the owning administration of model, if this model ratio is existing Model have less error rate, then with this model replace "current" model；

(5) if "current" model has sufficiently small error rate, or maximum iteration time has been reached, then algorithm terminates, Otherwise, (1), (2), (3), (4) are performed；

Such as Fig. 2, what parameterized model selected is affine transformation relationship, and now the green point covered is in two groups of data of A and B Meet the point of affine transformation relationship, be again intra-office point, other for point not in the know.In addition to calculating affine transformation matrix, RANSAC is also Such as Fig. 3 according to the parameterized model of the Demand Design of oneself oneself, can have found in a heap data and meet on same circle Most points and the parameter of this circle.

Fig. 2 show observation data A and observation data B in meet affine transformation relationship intra-office point drawn game exterior point minute Cloth, the intra-office point number now meeting affine relation reaches maximum, and can significantly find out that the number of point not in the know is more than The number of intra-office point, as above said introduction, this RANSAC algorithm and difference of method of least square just.

3 high frequency compensations

The computational methods of DSIFT are similar with SIFT, due to data set Ψ^*In image-region be adjusted, therefore solid Determine yardstick and the direction of SIFT, each pixel is calculated its SIFT and describes operator, then obtain DSIFT (dense sift) Descriptor

f_dj={ v_j,l_j,s,o} (6)

S, o represent changeless yardstick and direction, l respectively_j, represent f_djCentre coordinate in the picture, v_jIt is 128 The SIFT feature descriptor of dimension.

DSIFT matching algorithm, at image collection Ψ^*Each image MSER is carried out slip take block and calculate each image block DSIFT feature, to Pre-denoised image I_n ^*Carry out being similar to is slider-operated, uses the parameter that formula 5 provides as coupling The criterion of degree, the least matching degree of desired value is the highest, chooses the highest image block of matching degree and carries out high frequency compensation.AssumeThat a DSIFT coupling is right, then the image block after high frequency compensation

$I_{bc}=I_{b1}+\mathrm{\&gamma;}*(I_{b1}-I_{bx}^{*})---\left(7\right)$

Wherein I_b1Be original picture block to be compensated,For the noise-free picture block matched, γ is weight, takes here Empirical value 0.25.

Arthmetic statement

Sparse Denoising Algorithm framework based on the study of the similar sample in region is described as follows:

Input: multiple different artificial satellite threedimensional model set M, the artificial satellite image process of Noise: execution (1), (2)、(3)、(4)、(5)、(6)、(7)、(8)

(1), each the artificial satellite threedimensional model in M rotated, scale, project acquisition multi-pose multi-angle Two dimensional image set omega；

(2), to width image zooming-out SIFT-MSER feature bunch each in image collection Ω and put SIFT-MSER regional ensemble In Φ；

(3), extract the SIFT-MSER feature bunch of noise image and find similar in SIFT-MSER regional ensemble Φ All SIFT-MSER features bunch matched are put in similar sample set Ψ by characteristic block；

(4), SIFT-MSER feature bunch and noise image in similar sample set Ψ is estimated by RANSAC algorithm Affine transformation relationship between SIFT-MSER feature bunch, carries out inverse transformation to the image-region in data set Ψ, will match to MESR region is adjusted to the yardstick as noise MSER region, and the data after adjusting are stored in data set；

(5), data set Ψ^*In image-region carry out the image block of sliding window operation acquisition 8 × 8 as the sample of dictionary learning Example, obtained complete dictionary D by KSVD algorithm；

(6), use dictionary D and OMP algorithm that noise image carries out sparse reconstruct and obtain Pre-denoised image

(7), to data set Ψ^*Each image-region carries out sliding and takes block and calculate the DSIFT feature of each image block, puts Enter in DSIFT feature set Θ；

(8), for Pre-denoised imageCarrying out slides takes block and calculates the DSIFT feature of each image block, and at DSIFT Feature set Θ finds the DSFIT characteristic block mated most, utilizes and match DSFIT characteristic block to imageCarry out high frequency benefit Repay；

Output: image after denoising

Experimental result and analysis

For the effectiveness of verification algorithm, have chosen two different artificial satellite images in an experiment for being artificially generated Noise image, adds the white Gaussian noise of four kinds of intensity respectively, and noise intensity Sigma value is respectively 10,20,40,80.And with Portilla et al. proposes Bayes's least square-Gauss yardstick mixing Denoising Algorithm BLS-GSM (bayes least Squares-gaussian scale mixtures), Buades et al. non-local mean Denoising Algorithm NLM (Non-is proposed Local-Means) contrast test is done.

As Fig. 4 shows four kinds of algorithms denoising result on first group of test data, the longitudinal axis represents PSNR value, and transverse axis represents Noise intensity Sigma value.When Sigma takes 10, the present invention and BLSGSM algorithm denoising result are closer, and the most significantly larger than NLM calculates Method.Along with the denoising result of three kinds of algorithms of increase of noise intensity gradually reaches unanimity, three algorithms Sigma takes 80 when After denoising, PSNR is near 17dB.

Table 1 shows three kinds of methods concrete PSNR value on first group of test data, when in table, Sigma takes 10, this Bright obtain PSNR with BLSGSM of 35.8509dB, NLM algorithm compared with respectively high 0.2825dB, 4.2465dB, now NLM Algorithm PSNR is minimum.When Sigma takes 20, the present invention distinguishes height 0.6150dB, 2.4038dB compared with BLSGSM, NLM, now Noise intensity is moderate, relatively actual application, and three kinds of algorithm denoising result gaps are obvious.When Sigma takes 80, the present invention with BLSGSM, NLM have compared difference height 0.1008dB, 0.4569dB, and now three kinds of method denoising results are sufficiently close at PSNR.Always Body result is it appeared that the present invention is consistently higher than other two kinds of algorithms on PSNR.

1 first group of artificial satellite imaging experiments result PSNR of table

First group of artificial satellite image denoising result when noise intensity sigma=40, to artificial satellite image this The complete dictionary size of mistake of invention takes 64 × 256, and KSVD study dictionary iterations value is 10.The present invention is at red rectangle mark Being substantially better than other algorithms in visual effect in the region known, it is more clear than NLM algorithm to present invention obtains in circular antenna region Ground curved edge, present invention obtains than BLSGSM algorithm clearly at straight line and diagonal edge.

As Fig. 5 shows four kinds of algorithms denoising result on second group of test data, the longitudinal axis represents PSNR value, and transverse axis represents Noise intensity Sigma value.Consistent with the result of first group of experimental data, for noise strong relatively low time three kinds of algorithm denoising results poor Away from substantially, after noise intensity increases to certain value, three kinds of algorithm denoising results are sufficiently close on PSNR.

Table 2 shows three kinds of methods concrete PSNR value on second group of test data, in table Sigma value by 10 to 80 During increase, the present invention obtains the PSNR of 34.7928,29.4216,23.7856,17.7969, compared with other two kinds of algorithms all By higher PSNR, the PSNR after three kinds of algorithm denoisings sequence from low to high be followed successively by NLM, BLSGSM, the present invention propose Algorithm.Method is better than other two kinds of algorithms generally.

2 second groups of artificial satellite imaging experiments results PSNR of table

Second group of artificial satellite image denoising result, respectively show BLS-GSM Denoising Algorithm, NLM Denoising Algorithm and basis Invention Denoising Algorithm is to the artificial satellite image denoising effect when noise intensity sigma=40, the complete word of mistake of the present invention Allusion quotation size takes 64 × 256, and KSVD study dictionary iterations value is 10.The present invention is substantially better than in visual effect in region NLM algorithm.The Small object region that the present invention less enriches for edge obtains than BLSGSM algorithm clearly shape.

The denoising method that the present invention proposes not only obtains the Y-PSNR higher than BLS-GSM, NLM but also imitates in vision Also other two kinds of methods have been surmounted on fruit.Why the Denoising Algorithm framework that the present invention proposes can obtain good denoising effect Really, key reason is to change the study sample of complete dictionary, by selecting similar area to obtain study sample, affects The atomic structure of complete dictionary so that the atom in dictionary can preferably represent the live part of image, ignores noise section. By high frequency compensation further with the information of similar image, improve the signal to noise ratio of denoising image further.

The foregoing is only presently preferred embodiments of the present invention, not in order to limit the present invention, all essences in the present invention Any amendment, equivalent and the improvement etc. made within god and principle, should be included within the scope of the present invention.

Claims (6)

1. a sparse denoising method based on the study of the similar sample in region, it is characterised in that described based on the similar sample in region The sparse denoising method of study:

Image first by Noise carries out similar area based on SIFT-MSER feature bunch retrieval in image library, is obtaining The SIFT feature point coordinates of coupling is utilized to calculate corresponding affine transformation matrices after obtaining similar area；

The similar area conversion that will match to is to the direction identical with noise image and metric space；

Finally use the similar area after affine transformation as the sample of dictionary learning, improve dictionary relevant to noise image Property；Carry out high frequency compensation after completing based on new dictionary learning denoising again, improve the Edge texture characteristic information of image.

2. the sparse denoising method learnt based on the similar sample in region as claimed in claim 1, it is characterised in that similar area Acquisition specifically include:

The MSER feature extraction of image, after setting up sample image data set Ω, obtains number by MSER algorithm and SIFT algorithm Maximum stable extremal region and the scale invariant feature descriptor of intra-zone according to piece image every in collection Ω；Every piece image I presses specific structure after calculating MSER region and SIFT spy's feature and stores in file；Gray level image is carried out binaryzation Obtain the binaryzation sequence of a width figure；

Build the SIFT-MSER characteristic area bunch of image, first extract MSER region, in MSER region, then extract SIFT special Levying, M is the set in all MSER regions of image I, and S is the set of all SIFT feature of image I, and all characteristic areas bunch are remembered Make:

Λ=< { m₁,S₁},{m₂,S₂},…,{m_n,S_n}>；

M=< { m₁},{m₂},…,{m_n}>；

S=< { f₁},{f₂},…,{f_n}>；

Wherein, f_j={ v_j,l_j,s_j,o_jRepresent a SIFT feature descriptor, v_jIt is that this SIFT key point is at l_j=(x_j, y_j) 128 dimensional vector descriptors of position, s_jIt is the value of yardstick, o_jIt is the value in direction, S_j∈ S is a subset of S, table Show to fall the jth maximum stable extremal region m at image I_jInternal all SIFT feature describe operator；s_j={ b_j,l_j,a_j, b_jRepresent the region of MSER, b_jIt is the collection of pixels of this characteristic area, l_jIt is positional information, a_j,b_jRepresent oval respectively Major and minor axis；

Image local similar area based on SIFT-MSER feature bunch obtains, and closes setting up SIFT-MSER feature gatheringAfter, Syndrome setIn find out and noise image I_sIth feature bunch Λ_iThe feature bunch that ∈ Λ is similar.

3. the sparse denoising method learnt based on the similar sample in region as claimed in claim 2, it is characterised in that described similar Q is used in the acquisition in region₁, Q₂, Q_i-1, Q_iRepresent this series of mutually nested extremal region sequences:

$q\left(i\right)=\frac{\left|Q_{i+\mathrm{\&Delta;}}\right|-\left|Q_{i-\mathrm{\&Delta;}}\right|}{\left|Q_i\right|};$

Wherein i represents different threshold values, and Q is the set of a pixel, and what absolute value represented is this cardinality of a set, for extreme value The area in region.

4. the sparse denoising method learnt based on the similar sample in region as claimed in claim 2, it is characterised in that described basis The definition of Euclidean distance, the distance defined between two SIFT-MSER features bunch is:

$Dist({\mathrm{\&Lambda;}}_a,{\mathrm{\&Lambda;}}_b)=\sqrt{{\left(\frac{mean\left(M_a\right)-mean\left(M_b\right)}{255}\right)}^2+{\left(\frac{\left|S_a\right|+\left|S_b\right|}{2\&times;\left|match(S_a,S_b)\right|}\right)}^2};$

What absolute value represented is this cardinality of a set；mean(M_a) represent, match (S_a,S_b) represent S_aAt S_bIn have coupling element Set, obtain S_a,S_bThe SIFT feature descriptor of middle coupling；Each MSER region for noise image obtains C The region joined, then arrives the MSER regional ensemble Ψ of coupling.

5. the sparse denoising method learnt based on the similar sample in region as claimed in claim 1, it is characterised in that study sample Process use RANSAC algorithm, concrete grammar is as follows:

Input: by observation data D of noise severe contamination, parameterized mathematical model M, the unknown parameter m of parameterized model₁, m₂..., m_n；

Process: perform (1), (2), (3), (4) (5)

(1) selecting one group of data, at random is intra-office point, and other data are point not in the know；

(2) all parameters m of parameterized model, are calculated by intra-office point₁, m₂..., m_n；

(3), by parameter m obtained₁, m₂..., m_nGoing whether the point not in the know testing other meets parameterized model M, certain point is full Foot parameterized model M, moves into intra-office point by this point；

(4), use least square method, reappraise point in the owning administration of model；

(5), "current" model have sufficiently small error rate, or reached maximum iteration time, algorithm terminates, otherwise, perform (1)、(2)、(3)、(4)。

6. the sparse denoising method learnt based on the similar sample in region as claimed in claim 1, it is characterised in thatThat a DSIFT coupling is right, then the image block after high frequency compensation:

$I_{bc}=I_{b1}+\mathrm{\&gamma;}*(I_{b1}-I_{bx}^{*});$

Wherein I_b1Be original picture block to be compensated,For the noise-free picture block matched, γ is weight, takes empirical value here 0.25。