CN101951522B - Vector median filtering implementation method for eliminating pulse noises in color image - Google Patents
Vector median filtering implementation method for eliminating pulse noises in color image Download PDFInfo
- Publication number
- CN101951522B CN101951522B CN2010102972349A CN201010297234A CN101951522B CN 101951522 B CN101951522 B CN 101951522B CN 2010102972349 A CN2010102972349 A CN 2010102972349A CN 201010297234 A CN201010297234 A CN 201010297234A CN 101951522 B CN101951522 B CN 101951522B
- Authority
- CN
- China
- Prior art keywords
- vector
- pixel
- matrix
- vector distance
- filter window
- 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.)
- Expired - Fee Related
Links
Images
Landscapes
- Image Processing (AREA)
Abstract
The invention discloses a vector median filtering implementation method for eliminating pulse noises in a color image. When carrying out iterative computation on the vector distance matrix of the current pixel, the invention records and extracts part of computed vector distance values of the adjacent pixels in the vector distance matrix. The method can keep the same filtering effect as the original method, and can also effectively reduce the calculation time for filtering. The method enhances the operation efficiency of the classic vector filtering method and other expanded methods in the color image.
Description
Technical field
The present invention relates to a kind of implementation method of eliminating the vector median filtering of pulse noise in the coloured image.
Background technology
To the pulse noise that possibly occur in the collection and transmission of coloured image system; People such as Astola have proposed vector median filtering method VMF, referring to Astola J, and Haavisto P; Neuvo Y.VectorMedian Filters.In:Proc.IEEE.1990,78 (4): 678-689.This method can be eliminated such noise, also can avoid adopting the problem of the artificial colors that the standard medium filtering occurs.Because this method has become one of basic methods of coloured image denoising at present owing to have good robustness.Because edge and detailed information that classical vector medium filtering VMF can not well preserve original image, a lot of modified model vector median filtering methods are suggested successively.Be that classical vector median method VMF or a lot of modified model vector median filtering method ACWVDF, RSVMF, AVMF need calculate in the current filter window distance value between all vectors in filtering; In window, select output valve according to the different Rule of judgment of design more afterwards; Referring to Lukac R.Adaptive color image filtering based on center-weightedvector directional filters; Multidimensional Syst.Signal Processing; 2004; 15 (2): 169-196 and Lukac R.Adaptive color image filtering basedon center-weighted vector directional filters [J] .Syst.Signal Process.2004; 15:169-196 and Lukac R.Adaptive vector median filtering [J] .Pattern Recognition Letters, 2003,24 (12): I889-1899.The process of above vector median filtering method be iteration be every pair of pixel compute vectors distance matrix in each filter window that moves and the operation of selecting output vector.The step of distance value has caused the shortcoming that such filtering method efficient is not high, be difficult to satisfy the real-time operation demand between compute vectors.Distance calculating is the important channel of improving method efficient such as such vector median filtering between the interior vector of raising filter window.
Summary of the invention
Be to improve classical vector filtering method and the operation efficiency of its part extended method in the coloured image, the present invention provides a kind of Fast implementation of the vector median filtering to the vector distance matrix that needs to calculate each pixel.
Vector median filtering method (like VMF, ACWVDF, RSVMF, AVMF etc.) needs to calculate distance value between the vector in all windows.When the filter window size is m * n, filtering method will write size with distance value between all vectors and be the vector distance matrix of mn * mn.Filtering method is then selected output vector replacement original pixel value according to this matrix and filtering Rule of judgment method.The vector distance matrix of each filter window that the calculating that former vector median filtering method is an iteration is moved and the operation of selecting output vector.This method is being to need not the distance between all pixels in the calculation of filtered window for each pixel compute vectors distance matrix.Its value major part can come from the vector distance value of calculating in the middle of the corresponding vector distance matrix of contiguous a plurality of pixels, and current matrix vector distance value is from three types:
(1) the vector distance matrix of the calculating of the pixel in the colleague left side;
(2) being close to other pixel is stored in apart from storage matrix;
(3) recomputate.
Wherein the arrangement of vector numbering adopts numbering to begin from the upper left side in filter window, increases progressively earlier numbering from top to bottom more from left to right successively.
The corresponding vector distance matrix of each pixel wherein, its element satisfies along the diagonal symmetry, and the diagonal entry value is zero.
In its filtering iterative computation of full figure be in proper order: every successively from top to bottom row is handled, and in every row, adopts processing sequence from left to right.
After each pixel has been calculated the vector distance matrix, need the partial distance value is inserted in the corresponding vector storage matrix, so that other contiguous pixel is utilized this distance value.
Concrete technical scheme of the present invention is following:
When a kind of implementation method of eliminating the vector median filtering of pulse noise in the coloured image provided by the invention, filter window size were m * n, step was following:
Step 1: judge current pixel x
I, jWhether be edge pixel: if then adopt D
I, j(p, q)=|| x
p-x
q|| directly interior all vectors of calculation of filtered window are to x
p, x
qBetween distance value, forward step 3 afterwards again to; If not, then change step 2 over to;
Step 2: non-edge pixel x in the computed image
I, jCorresponding vector distance matrix D
I, j, the vector distance matrix D
I, jIn each element value be that each vector is to x
p, x
qBetween distance value adopt below wherein a kind of method obtain, wherein t1, t2 and t3 represent that different pieces of information comes Source Type:
T1: from the vector distance matrix D of the calculating of colleague's left side pixel
I-1, j
T2: with x
IjContiguous other other pixels are apart from the storage vector distance value among the storage matrix M;
T3: adopt || x
p-x
q|| recomputate the vector distance value;
Step 3: according to pixel x
I, jCorresponding vector distance matrix D
I, jCalculate replacement value y
I, j
Step 4: span is left matrix D
I, jThe segment vector distance value charge to apart from storage matrix M;
Step 5: if calculated last pixel of every row, then can change first pixel of next line over to and forward step 1 to, all calculate up to all pixels and finish.
Said size is the filter window of m * n, and vector is to x in filter window
p, x
qNumbering is arranged and is adopted numbering to begin from the upper left side, increases progressively numbering from top to bottom more from left to right successively earlier, and m and n are all odd number.
Said size is the filter window of m * n, according to each pixel filter window x '
L1... x '
MnIn coordinate (u, o), filter window is divided into subclass S1, S2, S3 and S4:
● S1={x '
11, x '
12..., x '
U, o..., x '
(n-1) (m-1) Wherein 1≤u≤(m-1) and 1≤o≤(n-1)
● S2={x '
M1..., x '
Mi... x '
M (n-1)1≤u≤(n-1) wherein
● S3={x '
1n..., x '
Jn... x '
(m-1) n1≤o≤(m-1) wherein
●S4＝{x′
mn}
According to vector to x
p, x
qSubclass under in filter window adopts the method for following table to judge that employing type t1, t2 or t3 come the compute vectors distance value.
Saidly comprised each pixel x apart from storage matrix M
I, jCorresponding M
I, j, when size is m * n filter window, storage x
I-m+2, j-n-1To x
I+m-2, jAll pixels and x in the rectangular extent
I, jThe segment vector distance value, charge to M
I, j(1,1) is to M
I, j(2m-3, n).
Said edge pixel is the pixel x in the full figure
I, jSet, satisfy:
Or
Beneficial effect of the present invention:
Be the operation efficiency of the classical vector filtering method in the raising coloured image and its part extended method, the present invention is openly to one type of fast method that needs the vector median filtering of all vector distances in the calculation of filtered window.When the value of the vector distance matrix of iterative computation current pixel, this invention record has also extracted its part from the vector distance value of in the vector distance matrix of neighborhood pixels, calculating.This new method can keep the filter effect identical with former method, and can effectively reduce the computing time of filtering.
Description of drawings
Fig. 1 is the pixel arrangement mode and the Data Source type of 5 * 5 filter windows.
Fig. 2 be 5 * 5 filter windows vector distance numerical value source type list type (p, q).
Fig. 3 is the current pixel x of 5 * 5 filter windows
I, jCorresponding to storage matrix M
I, j
Fig. 4 be 5 * 5 filter windows and current location for (100,100) corresponding apart from storage matrix M
100,100In the corresponding full figure coordinate position in each position.
Fig. 5 is the flow chart of vector median filtering Fast implementation.
Embodiment
Fig. 1 is the pixel arrangement mode and the Data Source type of 5 * 5 filter windows.
Each vector filtering method (like VMF, ACWVDF, RSVMF, AVMF etc.) is though its Rule of judgment has different with the vector distance calculation mode; But much all need distance between each vector in the calculation of filtered window; Promptly need iterative computation current location (i, j) the corresponding vector distance matrix D of pixel
I, j, when the filter window size was m * n, vector distance matrix D size was mn * mn, D
I, jIn each element definition following:
D
i，j(p，q)＝||x
p-x
q||
γ (x
p，x
q∈W)
W={x wherein
1..., x
Mn, || x
p-x
q||
γVector x in the expression W
pWith x
qThe norm distance, and satisfy: 1. || x
p-x
q||
γ=|| x
p-x
q||
γ2. || x
p-x
p||
γ=0.The moving right once of filter window iteration in filtering, promptly in the drawings from the full figure coordinate be (i, j-1) move to (i, in the time of j), the corresponding vector distance matrix D of new center pixel
I, jIn data need be from new calculating except that part, other can be from D
I, j-1With apart from obtaining among the storage matrix M, deposited each pixel that filter window moves through to deserved segment vector distance value apart from storage matrix M.D wherein
I, j(p, q) expression full figure coordinate is that (i j) is numbered distance between p and q vector in the filter window of pixel.Each vector numbering in the spectral window of design begins from the upper left side, and every from top to bottom leu increases progressively numbering.5 * 5 filter windows as shown in fig. 1 are in every lattice "/" preceding numeral is the number value of corresponding vector.And p
x, p
yAnd q
x, q
yBe respectively vector x
pWith x
qThe horizontal ordinate of window in filter window.
Be further to divide the vector distance source, may be partitioned into 4 son set at the filter window of m * n arbitrarily, according to each pixel filter window x '
11..., x '
MnIn coordinate (u, o), filter window is divided into subclass S1, S2, S3 and S4:
● S1={x '
11, x '
12..., x '
U, o..., x '
(n-1) (m-1) Wherein 1≤u≤(m-1) and 1≤o≤(n-1)
● S2={x '
M1..., x '
Mi... x '
M (n-1)1≤u≤(n-1) wherein
● S3={x '
1n..., x '
Jn... x '
(m-1) n1≤o≤(m-1) wherein
●S4＝{x′
mn}
5 * 5 filter windows as shown in Figure 1; Subclass S1 and S2 (being labeled as the pixel of */s1 and */s2 in the corresponding diagram) are preceding (n-1) row of m * n filter window, and subclass S3 and S4 (being labeled as the pixel of */s3 and */s3 in the corresponding diagram) are last row of filter window.
According to two vector x
pAnd x
qUnder subclass, definable vector distance value come Source Type, its type t1 refers to that this value derives from the vector distance matrix D
I, j-1Type t2 refers to that this value need recomputate; Type t3 refer to this value derive from store historical data apart from storage matrix M.
Fig. 2 has represented the type in every pair of vector distance value source in the filter window in the present invention, for example as vector x
p∈ S1 ∪ S2 and x
q∈ S1 ∪ S2, distance value is directed to D
I, j-1, promptly type (p, q)=' t1 '; And for example as vector x
p∈ S3 and x
qDuring ∈ S3, distance value is from apart from obtaining the storage matrix M, promptly type (p, q)=' t3 '.Consider the character (symmetry) of vector distance norm, type list also is about the diagonal symmetry.
According to above definition, every couple of vector distance D in the filter window
I, j(p, q) computational methods are following:
type(p，q)＝t1
When type (p, q)=t2
type(p，q)＝t3
Apart from storage matrix M is one 4 dimension matrix.Wherein preceding 2 tie up the position of pixel in full figure of the correspondence of expression, the relative coordinate of certain pixel among (the 2m-3) * n that dimension is represented and the respective pixel left is contiguous of back 2.If the full figure coordinate of current pixel is that (i j), has then stored the full figure coordinate and has been [i-m+2..i+m-2], and pixel and full figure coordinate in [j-n+1...j] scope are (i, the vector distance value of pixel j).Like Fig. 3, if the filter window size is 5 * 5 o'clock, what then each pixel storage was corresponding is 7 * 5 apart from vector value size among the storage matrix M, wherein the current pixel in the corresponding full figure in (4,5) position.If the full figure coordinate of current pixel be (100,100), then in its corresponding matrix M, stored the full figure coordinate range and be ([97 ... 103], [96 ... 100] 35 pixels) and coordinate are distance value between (comprising self) vector of (100,100) pixel.Coordinate is full figure coordinate such as Fig. 4 apart from corresponding diagram 3 each position among the storage matrix M of (100,100) pixel.For example Fig. 3 and Fig. 4 can know, full figure coordinate (100,100) pixel apart from storage matrix M in position (1,5) write down the vector distance value of full figure coordinate (97,100) and (100,100) pixel.
Pixel (i, j) the vector distance matrix D of correspondence according to definition
I, jComputational methods can know, when calculating the full figure coordinate and be the vector distance matrix of (98,98), can calculate for the first time obtain the full figure coordinate be ([97 ... 100], [96 ... 100]) interior 20 pixels of scope and full figure coordinate are the distance value of (100,100) pixel.If with these values count the full figure coordinate for the respective distances storage matrix coordinate of (100,100) pixel be ([1 ... 4], [1 ... 5]).When calculating full figure coordinate was the vector distance matrix of (99,98) pixel, the coordinate that then wherein needs was (100; 100) vector and scope be ([97 ..., 100]; [96 ..., 100]) 20 vector distance values in the scope then need not to recomputate; As long as from coordinate is getting final product apart from obtaining among the storage matrix M of (100,100) pixel.
Can know that according to above description distance comprises the n+m-2 between (nm-1) between each element of S set 4 and S1 ∪ S2 ∪ S3 inferior and S2 and the S3, nm+n+m-3 time altogether between the vector that this Fast implementation need recomputate.It is mn (mn-1)/2 that distance vector matrix in the former vector median filtering method needs calculation times.
Successively each pixel is carried out filtering through order mobile filter window.Preceding method has been considered the calculating of the vector distance matrix of the non-edge pixel of full figure, considers that also the vector distance matrix of edge of image pixel must obtain mode.And then this fast the flow process of implementation method be Fig. 5.
Step S601: judge that whether current location is last location of pixels that surpasses image, is then to quit a program, otherwise continues;
Step S602: judging whether current location is edge pixel, is the account form that then adopts the conventional vector distance matrix;
Step S603: traditional approach calculates the vector distance matrix of current pixel, and for example under 5 * 5 filter window, the number of times of the vector distance that then need calculate is 300 times, changes step S609 over to;
Step S604: if current pixel is not edge pixel, then can promptly go up from the left side vector distance matrix acquisition type that a pixel calculated (p, q)=' the segment vector distance value of t1 ';
Step S605: need recomputate type (p, q)=' another part distance value in the new vector distance matrix of t2 ' then;
Step S606: for type (p, q)=' the remaining part of t3 ', need from the vector distance storage matrix M of neighborhood pixels, obtain;
Step S607: the segment vector distance that calculates is filled among the corresponding vector distance storage matrix M of this pixel;
Step S608: the symmetry and the distance matrix diagonal that calculate according to vector distance are zero characteristic, calculate complete vector distance matrix;
Step S609:, calculate condition output vector y separately according to the different filtering method
Ij
Step S610: if current pixel is last pixel of every row, a pixel then moves right; Get back to step S601.
Claims (4)
1. implementation method of eliminating the vector median filtering of pulse noise in the coloured image is characterized in that: the filter window size is during for m * n, and step is following:
Step 1: judge current pixel x
I, jWhether is edge pixel, wherein i, j are respectively horizontal stroke, the ordinate of current pixel in image; If then adopt D
I, j(p, q)=|| x
p-x
q|| directly interior all vectors of calculation of filtered window are to x
p, x
qBetween distance value, forward step 3 afterwards again to; If not, then change step 2 over to, edge pixel is to satisfy in the full figure
Or
Pixel;
Step 2: non-edge pixel x in the computed image
I, jCorresponding vector distance matrix D
I, j, the vector distance matrix D
I, jIn each element value be that each vector is to x
p, x
qBetween distance value D
I, j(p q) adopts wherein a kind of method of t1, t2 or t3 to obtain, and wherein t1, t2 and t3 represent that different pieces of information comes Source Type:
T1: from the vector distance matrix D of the calculating of colleague's left side pixel
I-1, j
T2: with x
I, jContiguous other pixels are apart from the storage vector distance value among the storage matrix M;
T3: adopt || x
p-x
q|| recomputate the vector distance value;
Step 3: according to pixel x
I, jCorresponding vector distance matrix D
I, jCalculate replacement value y
I, j
Step 4: vector distance matrix D
I, jThe segment vector distance value charge to apart from storage matrix M;
Step 5: move right to next pixel, if calculated last pixel of every row, then can change first pixel of next line over to and forward step 1 to, all calculate up to all pixels and finish.
2. method according to claim 1 is characterized in that: said size is the filter window of m * n, and vector is to x in filter window
p, x
qNumbering is arranged and is adopted numbering to begin from the upper left side, increases progressively numbering from top to bottom more from left to right successively earlier, and m and n are all odd number.
3. method according to claim 1 is characterized in that: said size is the filter window of m * n, according to each pixel filter window x '
11..., x '
MnIn coordinate (u, o), filter window is divided into subclass S1, S2, S3 and S4:
● S1={x '
11, x '
12..., x '
Uo..., x '
(n-1) (m-1)Wherein 1≤u≤(m-1) and 1≤o≤(n-1)
● S2={x '
M1..., x '
Mi... x '
M (n-1)1≤u≤(n-1) wherein
● S3={x '
1n..., x '
Jn... x '
(m-1) n1≤o≤(m-1) wherein
●S4＝{x′
mn}
According to vector to x
p, x
qSubclass under in filter window adopts the method for following table to judge that employing type t1, t2 or t3 come the compute vectors distance value.
4. method according to claim 1 is characterized in that: saidly comprised each pixel x apart from storage matrix M
I, jCorresponding M
I, j, when size is m * n filter window, M
I, jBe storage x
I-m+2, j-n+1To x
I+m-2, jAll pixels and x in the rectangular extent
I, jThe matrix of segment vector distance value, charge to M
I, j(1,1) is to M
I, j(2m-3, n).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2010102972349A CN101951522B (en) | 2010-09-28 | 2010-09-28 | Vector median filtering implementation method for eliminating pulse noises in color image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2010102972349A CN101951522B (en) | 2010-09-28 | 2010-09-28 | Vector median filtering implementation method for eliminating pulse noises in color image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101951522A CN101951522A (en) | 2011-01-19 |
| CN101951522B true CN101951522B (en) | 2012-05-23 |
Family
ID=43454860
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2010102972349A Expired - Fee Related CN101951522B (en) | 2010-09-28 | 2010-09-28 | Vector median filtering implementation method for eliminating pulse noises in color image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101951522B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103455627B (en) * | 2013-09-18 | 2017-06-06 | 国家电网公司 | A kind of data sharing method and device |
| CN104597499B (en) * | 2013-10-31 | 2017-02-08 | 中国石油天然气集团公司 | Adjacent shot interference suppressing method and device for independent simultaneous sweeping seismic data of seismic vibrators |
| CN104849760B (en) * | 2015-06-02 | 2017-07-14 | 中国矿业大学(北京) | The Vector Median Filtering Method converted based on Tau p |
| CN105701779B (en) * | 2016-01-11 | 2019-02-05 | 中国科学院自动化研究所 | Color image median filter method and system based on fuzzy comprehensive evoluation |
| CN106327436A (en) * | 2016-08-10 | 2017-01-11 | 广东工业大学 | Filtering method and device based on multichannel noise detection |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6229578B1 (en) * | 1997-12-08 | 2001-05-08 | Intel Corporation | Edge-detection based noise removal algorithm |
| CN101340594A (en) * | 2007-07-06 | 2009-01-07 | 北京大学软件与微电子学院 | Color image pulse noise eliminating method and system based on vector outer product and median filtering |
-
2010
- 2010-09-28 CN CN2010102972349A patent/CN101951522B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6229578B1 (en) * | 1997-12-08 | 2001-05-08 | Intel Corporation | Edge-detection based noise removal algorithm |
| CN101340594A (en) * | 2007-07-06 | 2009-01-07 | 北京大学软件与微电子学院 | Color image pulse noise eliminating method and system based on vector outer product and median filtering |
Non-Patent Citations (2)
| Title |
|---|
| 钟灵等.双阈值开关型矢量中值滤波.《计算机应用研究》.2010,第27卷(第6期),2367-2369. * |
| 韩金玉等.基于噪声特征和矢量中值滤波的彩色图像去噪算法.《计算机应用》.2009,第29卷(第9期),2418-2419,2423. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN101951522A (en) | 2011-01-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110119780B (en) | Hyper-spectral image super-resolution reconstruction method based on generation countermeasure network | |
| Deng et al. | Wavelet domain style transfer for an effective perception-distortion tradeoff in single image super-resolution | |
| CN101951522B (en) | Vector median filtering implementation method for eliminating pulse noises in color image | |
| CN109086773B (en) | Fault plane identification method based on full convolution neural network | |
| CN110634147B (en) | Image matting method based on bilateral guide up-sampling | |
| CN105741252A (en) | Sparse representation and dictionary learning-based video image layered reconstruction method | |
| CN110232394A (en) | A kind of multi-scale image semantic segmentation method | |
| CN103632359B (en) | A kind of video super-resolution disposal route | |
| CN101930604B (en) | Infusion method of full-color image and multi-spectral image based on low-frequency correlation analysis | |
| CN101609554B (en) | Gray-scale template matching method based on gray-scale statistical characteristics | |
| CN109919883B (en) | Traffic video data acquisition method based on gray level conversion | |
| CN102402784B (en) | Human face image super-resolution method based on nearest feature line manifold learning | |
| CN103413286A (en) | United reestablishing method of high dynamic range and high-definition pictures based on learning | |
| CN102800094A (en) | Fast color image segmentation method | |
| CN104036479A (en) | Multi-focus image fusion method based on non-negative matrix factorization | |
| CN111383200B (en) | CFA image demosaicing method based on generated antagonistic neural network | |
| CN101635048A (en) | Super-resolution processing method of face image integrating global feature with local information | |
| CN102542296A (en) | Method for extracting image characteristics by multivariate gray model-based bi-dimensional empirical mode decomposition | |
| CN106097252A (en) | High spectrum image superpixel segmentation method based on figure Graph model | |
| CN105701770A (en) | Face super-resolution processing method and system based on context linear model | |
| CN106600533A (en) | Single-image super-resolution reconstruction method | |
| CN116468645A (en) | Antagonistic hyperspectral multispectral remote sensing fusion method | |
| CN104504672A (en) | NormLV feature based low-rank sparse neighborhood-embedding super-resolution method | |
| CN105701515A (en) | Face super-resolution processing method and system based on double-layer manifold constraint | |
| CN105335930A (en) | Edge data driven robustness-based face super-resolution processing method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| C17 | Cessation of patent right | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20120523 Termination date: 20120928 |