CN102609914A - Signal-correlated noise estimating method for image sensor - Google Patents
Signal-correlated noise estimating method for image sensor Download PDFInfo
- Publication number
- CN102609914A CN102609914A CN201210014261XA CN201210014261A CN102609914A CN 102609914 A CN102609914 A CN 102609914A CN 201210014261X A CN201210014261X A CN 201210014261XA CN 201210014261 A CN201210014261 A CN 201210014261A CN 102609914 A CN102609914 A CN 102609914A
- Authority
- CN
- China
- Prior art keywords
- noise
- image
- expression
- smooth
- sample point
- 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.)
- Granted
Links
Images
Landscapes
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a computer image processing and visual detecting genecology. In order to provide a method which can exactly estimate the noise function of different CCD (charge coupled device) images and further can not increase the computing complexity, the technical scheme adopted by the invention is as follows: a signal-correlated noise estimating method for an image sensor comprises the following steps of: searching smooth blocks of an image by adopting a high pass operator-based image structure analyzer; sampling the blocks according to the arrangement mode of CFA (colorful filter array), to obtain a set which corresponds to each gray value, and computing the mean value and variance of the set, to obtain a set of noise estimating sample points; according the smoothness degree of the set of the image block of each sample point, computing the reliability of the sample point; and finally in combination with the sample points, the reliability of the sample points and the base obtained by using PCA (principal component analysis) in advance, rebuilding a noise horizontal curve by a sparse representation technology. The method is mainly used for processing the image.
Description
Technical field
The present invention relates to Computer Image Processing, vision detection technology, especially relate to the signal dependent noise method of estimation of imageing sensor.
Background technology
In recent years, be accompanied by digital camera, have the popularizing of mobile phone of camera, the mankind have got into whole people's digital age, and have become more and more important naturally as the imageing sensor of the core devices of digital camera.The imageing sensor that current quilt generally adopts comprises two kinds: CCD (Charge-Coupled Device) sensor and cmos sensor.Wherein beam coupling device (CCD) with its have self-scanning, high resolving power, advantage such as output noise is low, dynamic range is big, quantum efficiency is high, charge transfer effciency is high, spectral response range is wide, geometrical stability is good, be widely used in fields such as shooting, IMAQ, scanner and commercial measurement at present.
But the CCD noise comprises input and output at work, in opto-electronic conversion and the charge transfer process, all can produce noise.This comprising fixed pattern noise, dark current noise, shot noise, remove mosaic noise and quantizing noise etc.Add the nonlinear interaction of camera response function CRF (Camera Response Function), causing the CCD noise has not been simple white Gaussian noise, but a kind of noise SDN (Signal-Dependent Noise) that depends critically upon signal intensity.Traditional stack white Gaussian noise AWGN (Additive White Gaussian Noise) model is inapplicable fully for the CCD imaging system.
At present, the method for estimation for the sort signal correlation noise mainly comprises three major types.One type of Noise Estimation that is based on the smooth image piece.This method mainly is through calculating mean absolute deviation MAD (Mean Absolute Deviate), seeking smooth in the entire image.Through calculating the average and the variance of smooth image piece, obtain the sampling point of noise curve, utilizing approach based on linear interpolation to obtain noise curve.This method simple and fast can estimate the noise level of image to a certain extent, but because its adaptivity is poor, and the curve that obtains is very coarse, and relevant with the content of image, less stable.Second type of method of estimation that is based on mathematic(al) manipulation.These class methods mainly are to utilize linear transformation, and for example wavelet transformation or dct transform etc. carry out denoising to image, utilize the residual error of former noise image and estimated image to come the reconstruct curve then.It is last type that the effect of these class methods is better than, and the curve that reconstructs is better, but the complexity of calculating is higher, and for the image of many textures or little noise, is easy to occur situation about estimating.Simultaneously, how to be a difficult problem in the computer vision equally accurately to the SDN image denoising.The 3rd type of method that is based on the probability derivation come the estimating noise curve.This method was utilized piecewise smooth iconic model before this; Estimate the upper bound of noise level function NLF (Noise Level Function); Utilize the correlativity of three passages of coloured image again,, calculate noise curve according to Bayes's maximum posteriori criterion.These class methods can realize the estimation to the CCD noise curve preferably, but same computation complexity is higher.Simultaneously, because this method is based on image segmentation, for the image that part is not easy to do and cuts apart, effect is not fine, and for the less ccd image of noise, the problem of estimating can occur yet.
Summary of the invention
The present invention is intended to solve the deficiency that overcomes prior art, and a kind of signal dependent noise method of estimation of imageing sensor is provided, and can accurately estimate the noise function of different ccd images, can't increase the complexity of calculating simultaneously.For achieving the above object; The technical scheme that the present invention takes is; A kind of signal dependent noise method of estimation of CCD image sensor; CCD is the prefix abbreviation of beam coupling device Charge-Coupled Device, it is characterized in that, comprises the steps: to adopt based on the picture structure analyzer of high pass operator to seek smooth of image; Then these pieces are sampled according to the arrangement mode of color filter Matrix C FA; CFA is the prefix abbreviation of Color Filter Array; Obtain a corresponding set of each gray-scale value, calculate the average and the variance of this set, obtain one group of Noise Estimation sample point; According to the smooth degree that obtains that set of diagrams picture piece of each sample point, calculate the confidence level of this sample point simultaneously; The horizontal function NLF of calculating noise function library is done principal component analysis (PCA) to it, the proper vector set that calculates, as the dictionary of rarefaction representation; Combine sample point, sample point confidence level at last and utilize principal component analysis (PCA) in advance, the technology through rarefaction representation reconstructs the noise level curve.
Employing is sought smooth of image based on the picture structure analyzer of high pass operator and further is refined as:
11) the picture structure analyzer is made up of eight high pass operators, and image is smooth more, and the value of calculating is more little;
12) in the G passage, get image block
and the step-length of getting piece; Be to follow the big or small W of image block to interrelate; The number that should guarantee piece can not reduce calculated amount very little again as far as possible;
Wherein, (1)
Here
Centre coordinate is positioned at (i, image block j), I in the expression G passage
(G)(k, l) coordinate is (k, pixel value l), W in the expression G passage
IjThe presentation video piece
The coordinate range of pixel;
13) according to the template M of picture structure analyzer
n, n=1,2......7,8, computing center's coordinate be positioned at (i, the smooth degree of image block j), here
The image block of expression and the template of structure analyzer are made convolution;
14) average of computing block is as the grey level of this image block;
According to the smooth degree of piece, calculate and to define the whether smooth threshold curve of image block, find out all smooth image set of blocks in three passages of entire image;
21) to all images piece of G passage,, carry out statistical average, calculate the smooth degree curve of image on whole gray space according to the grey level of piece;
Wherein,
Here, S
IjThe index set of the similar image block of expression grey level, the grey level of L presentation video piece, ε represent the gray-scale deviation that allows, | S
Ij| the expression S set
IjThe number of middle element;
22) according to the size of current block, select a suitable parameters λ, just obtained the judgement image block
Whether smooth threshold tau
Ij
The corresponding smooth degree of grey level
in
expression entire image is here calculated by formula (4);
23) according to threshold curve; Judge whether each G channel image piece is smooth, for being considered to smooth image block, the R passage of same position and the image block in the B passage; Also be considered to smooth, so just obtained the set of all smooth image pieces of entire image;
Arrangement mode according to color filter Matrix C FA is sampled, and obtains a corresponding set of each gray-scale value, calculates the average and the variance of this set, obtains one group of Noise Estimation sample point, is specially:
31) according to the arrangement mode of the colour filter Matrix C FA of CCD camera, to extract out respectively in each passage without the pixel of interpolation, the gray-scale value level according to pixel place image block constitutes a new set, remembers that the sample mode of three passages of CFA is M
(R), M
(G), M
(B), then the new set that obtains to gray-scale value L is:
Wherein,
And
The e here representes the sample mode dot product of each channel image piece and CFA, samples out without the pixel of interpolation;
32) calculate the average and the variance of this set, just obtained the sample point
of a Noise Estimation
The I here
LThe pixel level that expression is calculated through the corresponding collection of pixels of grey level L,
The horizontal I of remarked pixel
LCorresponding noise level.
According to the smooth degree that obtains that set of diagrams picture piece of each sample point, calculate the confidence level of this sample point simultaneously, be specially:
41) same images of positions piece in each passage is endowed the confidence level that it equates, promptly
The ω here
IjExpression is centered close to (i, the confidence level of image block j), the rate of decay of h control confidence level;
42) obtain the collection of pixels C of gray-scale value L
LCorresponding confidence level set
Wherein,
43) this confidence level set is asked average, just obtained the corresponding confidence level of this sample point;
The horizontal function NLF of calculating noise function library is done principal component analysis (PCA) to it, and the proper vector set that calculates, the dictionary as rarefaction representation is specially:
51) in order to estimate the noise level function NLF of various types of cameras under different noise intensities, the CCD noise model of structure is:
I=f(L
I+n
s+n
c)+n
q (14)
Wherein, I representes the noise image that actual observation is arrived, L
IThe intensity of illumination that presentation video is desirable, n
sExpression depends on intensity of illumination L
INoise, n
cExpression does not rely on and intensity of illumination L
INoise, n
qExpression quantizing noise, and f () expression camera response function CRF;
52) when the input one width of cloth from 0 to 255 gradually change standard picture the time, obtain a width of cloth CCD noise image, utilize following formula, just can obtain a NLF curve;
The f here representes camera response function, σ
sExpression depends on intensity of illumination L
IThe standard deviation of noise, σ
cExpression does not rely on intensity of illumination L
IThe standard deviation of noise, I
N(g) building-up process of expression noise, E [g] expression is asked expectation to the operation result of the inside;
53) according to the situation of the CCD camera of real world, choose the suitable parameters scope, work as f, n
s, n
cWhen changing successively, just obtained a NLF function library;
54) this NLF function library is done principal component analysis (PCA), obtain the eigenwert and its characteristic of correspondence vector of these curves, the dictionary D of the set of proper vector as rarefaction representation.
Combine sample point, sample point confidence level at last and utilize principal component analysis (PCA) in advance, the technology through rarefaction representation reconstructs the noise level curve:
61) sample point is designated as y ∈ R
M * 1, m<<256, the R here
M * 1What represent is the set of m on the real number space * 1 dimensional vector, and original noise level signal is designated as x ∈ R
256 * 1, the relational expression below x and y satisfy
y=Фx (16)
Ф ∈ R wherein
M * 256, every row has only one 1 element, place row corresponding the gray level at y place, be equivalent to a sampling matrix;
62) the rarefaction representation theory shows, if signal x has rarefaction representation under one group of base D, the reconstruct of signal x just can be converted into the sparse solution of finding the solution the consistent equation of observation so, promptly
s.t.||W(y-ФDα)||
2<ε (17)
Wherein, || g||
1With || g||
2Represent L respectively
1Norm and L
2Norm, s.t. representes to make satisfied, and α waits to ask the rarefaction representation coefficient of signal x under dictionary D, and W representes the confidence level of each sample, and ε is a very little number, the curve of expression and the approximation ratio of sample.Utilizing second order cone planning is that second-order cone programming method is found the solution (17) formula, the rarefaction representation
of noise level function curve x under D that just can obtain CCD then reconstruct CCD noise level function curve is
Technical characterstic of the present invention and effect:
1, method is simple, and the complexity of algorithm is low, is easy to realize;
2, among the present invention, when seeking the smooth image piece, be not to judge, but calculate with a kind of picture structure analyzer according to the variance of image block;
3, when judging that image block is whether smooth; Not to be used as threshold value with a simple value; But according to the smooth degree curve of entire image at whole gray space; Calculate a threshold value, can avoid the estimation of owing of crossing estimation and big noise region of little noise region to a great extent along with the noise variation;
4, the noise image among the present invention all is a coloured image, when we choose the smooth image piece, is not to calculate by passage, but only calculates the G passage, thinks that the image of same position has identical smooth property in three passages.So both can reduce the influence of demosaic algorithm, more can reduce calculated amount;
5, in the sample point that utilizes smooth image piece calculating noise to estimate; Because the influence of color filter Matrix C FA sampling; Be not directly to estimate, but according to the arrangement mode of CFA with the piece variance, in each piece of sampling out without the pixel of interpolation; Pixel level according to image block gathers together, and average and the variance of calculating this set are as a sample point;
6, the confidence level of each sample point is different, and this depends on the smooth degree of that set of diagrams picture piece that calculates this sample point, and image block is smooth more, and the sample point that obtains is reliable more;
7, according to the imaging process of CCD camera, make up noise model, according to the type of the CCD camera of real world, select suitable parameters then, calculate a function library, after analyzing through PCA, for curve Reconstruction provides one group of base;
8, utilize the confidence level of sample point and sample point, utilize the technology of rarefaction representation, just can well reconstruct the noise level function curve of ccd image;
The new method of the CCD Noise Estimation that 9, proposes among the present invention has favorable expansibility and practicality, can improve effect for other algorithm estimated parameter of computer vision field.
Description of drawings
Block diagram of using among the present invention and experimental result are explained as follows:
Fig. 1 is based on the block diagram of the CCD Noise Estimation of smooth block search and rarefaction representation;
The template (5x5) of Fig. 2 picture structure analyzer is used for weighing the smooth degree of an image block;
Fig. 3 simulates the synthetic block diagram of CCD noise, is used for the function library of the horizontal function of generted noise (NLF);
The test pattern design sketch that Fig. 4 constructs in order to generate the NLF function library;
The noise image of Fig. 5 test pattern goldhill, noise parameter are got f (●)=CRF (60), σ
s=0.10, σ
c=0.04;
The Noise Estimation sample point of Fig. 6 goldhill image, the sample point that wherein black circle expression estimates, stain are represented the CCD noise of simulating;
The noise level curve that Fig. 7 utilizes the goldhill image reconstruction to come out, wherein dotted line is represented the curve that reconstruct is come out, solid line is represented desirable CCD noise curve;
The noise image of Fig. 8 test pattern boats, noise parameter are got f (●)=CRF (60), σ equally
s=0.10, σ
c=0.04;
The Noise Estimation sample point of Fig. 9 boats image, the sample point that wherein black circle expression estimates, stain are represented the CCD noise of simulating;
The noise level curve that Figure 10 utilizes the reconstruct of boats image to come out, wherein dotted line is represented the curve that reconstruct is come out, solid line is represented desirable CCD noise curve;
The noise image of Figure 11 test pattern pen, noise parameter are got f (●)=CRF (60), σ
s=0.06, σ
c=0.02;
The Noise Estimation sample point of Figure 12 pen image, the sample point that wherein black circle expression estimates, stain are represented the CCD noise of simulating;
The noise level curve that Figure 13 utilizes the reconstruct of pen image to come out, wherein dotted line is represented the curve that reconstruct is come out, solid line is represented desirable CCD noise curve;
The noise image of Figure 14 test pattern pepper, noise parameter are got f (●)=CRF (120), σ
s=0.06, σ
c=0.02;
The Noise Estimation sample point of Figure 15 pepper image, the sample point that wherein black circle expression estimates, stain are represented the CCD noise of simulating;
The noise level curve that Figure 16 utilizes the reconstruct of pepper image to come out, wherein dotted line is represented the curve that reconstruct is come out, solid line is represented desirable CCD noise curve;
Embodiment
The technical matters that the present invention will solve is to propose a kind ofly to estimate noise sample through the smooth image piece, utilizes the confidence level of these noise sample values and sample value again, utilizes the technology of rarefaction representation to reconstruct the method for noise level function (NLF) curve.This method not only wants accurately to estimate the noise function of different ccd images, can't increase the complexity of calculating simultaneously.
Technical scheme of the present invention is:
Adopt a kind of method, smooth of seeking image based on picture structure rather than traditional calculating image block variance.Then these pieces are sampled according to the arrangement mode of color filter Matrix C FA, obtain a corresponding set of each gray-scale value, calculate the average and the variance of this set, just obtain one group of Noise Estimation sample point.According to the smooth degree that obtains that set of diagrams picture piece of each sample point, calculate the confidence level of this sample point simultaneously.Combine sample point, sample point confidence level at last and utilize PCA to analyze in advance the base obtain, the technology through rarefaction representation reconstructs the noise level curve.The method that experiment showed, us can well estimate the CCD noise level, and complexity is lower.Concrete grammar may further comprise the steps:
1) utilizes the picture structure analyzer, choose suitable step-length, calculate the smooth degree of image block in the CCD noise image G passage, and the average of computing block is as the grey level of this piece according to the size of image;
11) the picture structure analyzer is as shown in Figure 2, is made up of eight high pass operators, and image is smooth more, and the value of calculating is more little;
12) in the G passage, get image block
because the G passage receives the influence of demosaic algorithm minimum.And the step-length of getting piece is to follow the big or small W of image block to interrelate, and should guarantee that the number of piece can not reduce calculated amount very little again as far as possible;
Wherein, (1)
13) according to the template M of picture structure analyzer
n, n=1,2......7,8, calculate the smooth degree of each piece;
14) average of computing block is as the grey level of this image block;
2) according to 1) in the smooth degree of piece, calculate and define the whether smooth threshold curve of image block, find out all smooth image set of blocks in three passages of entire image;
21) to all images piece of G passage,, carry out statistical average, calculate the smooth degree curve of image on whole gray space according to the grey level of piece, wherein, S
IjThe index set of the similar image block of expression grey level, ε representes the gray-scale deviation that allows, | S
Ij| the expression S set
IjThe number of middle element;
Wherein,
22) according to the size of current block, select a suitable parameters λ, just obtained the judgement image block
Whether smooth threshold tau
Ij
23) according to threshold curve; Judge whether each G channel image piece is smooth, for being considered to smooth image block, the R passage of same position and the image block in the B passage; Also be considered to smooth, so just obtained the set of all smooth image pieces of entire image;
3), sample the sample point that calculating noise is estimated according to the arrangement mode of the colour filter Matrix C FA of CCD camera to each smooth image piece;
31) because the influence of demosaic; Directly the variance of computed image piece is as noise variance; But according to the arrangement mode of the colour filter Matrix C FA of CCD camera; Extract in each passage the pixel without interpolation respectively out, the gray-scale value level according to pixel place image block constitutes a new set.The sample mode of three passages of note CFA is M
(R), M
(G), M
(B), then the new set that obtains to gray-scale value L does
Wherein,
And
32) calculate the average and the variance of this set, just obtained the sample point
of a Noise Estimation
4), calculate the degree of confidence of this sample point according to the smooth degree that obtains all pieces of a sample point;
41) confidence level of each sample point is to be closely connected with those smooth image pieces that calculate this sample point.Because the smooth degree of R passage and B passage is to weigh through the G channel image piece of same position.So same images of positions piece in each passage is endowed the confidence level that it equates, promptly
The rate of decay of the h control confidence level here;
42) be similar to the process of sample estimates point in (3), can obtain the collection of pixels C of gray-scale value L too
LCorresponding confidence level set
Wherein,
43) this confidence level set is asked average, just obtained the corresponding confidence level of this sample point;
5) calculate the NLF function library, it is done principal component analysis (PCA), extract bigger several characteristic vector, as the base of rarefaction representation;
51) in order to estimate the noise level function (NLF) of various types of cameras under different noise intensities, the CCD noise model of structure is:
I=f(L+n
s+n
c)+n
q (14)
Wherein, I representes the noise image that actual observation is arrived, the intensity of illumination that the L presentation video is desirable, n
sExpression depends on the noise of intensity of illumination L, n
cExpression does not rely on the noise with intensity of illumination L, n
qThe expression quantizing noise because value is smaller, can be ignored it.And the synthetic process of CCD noise of f () expression, as shown in Figure 3, the CRF here representes camera response function (Camera Response Function).
52) when the input one width of cloth from 0 to 255 gradually change standard picture (like Fig. 4) time, we just can obtain a width of cloth CCD noise image.Utilize following formula, just can obtain a NLF curve;
53) according to the situation of the CCD camera of real world, choose the suitable parameters scope.Work as f, n
s, n
cWhen changing successively, just obtained a NLF function library;
54) this NLF function library is done principal component analysis (PCA), obtain the proper vector of these curves, get maximum preceding 10 proper vectors of institute's character pair value and constituted the dictionary D that is used for rarefaction representation;
6) the Noise Estimation sample point that utilizes the front to obtain, and the degree of confidence of sample point utilize the method reconstruct NLF curve of rarefaction representation.
61) sample point is designated as y ∈ R
M * 1(m<<256), original noise level signal is designated as x ∈ R
256 * 1, the relational expression below x and y satisfy
y=Фx (16)
Ф ∈ R wherein
M * 256, every row has only one 1 element, place row corresponding the gray level at y place, be equivalent to a sampling matrix.
62) the rarefaction representation theory shows, if signal x has rarefaction representation under one group of base D, the reconstruct of signal x just can be converted into the sparse solution of finding the solution the consistent equation of observation so, promptly
s.t.||W·(y-ФDα)||
2<ε (17)
Wherein α waits to ask the rarefaction representation coefficient of signal x under dictionary D, and W representes the confidence level of each sample, and ε is a very little number, the curve of expression and the approximation ratio of sample.Utilize second order cone planning (second-order cone programming) method to find the solution (17) formula, the rarefaction representation
of noise level function curve x under D that just can obtain CCD then reconstruct CCD noise level function curve is
Elaborate below in conjunction with instance and accompanying drawing new method to the imageing sensor Noise Estimation based on smooth estimation and rarefaction representation technology of the present invention.
Whether the method for the new CCD Noise Estimation that the present invention proposes at first is to utilize a kind of picture structure analyzer based on the high pass operator, rather than the variance of image block, come decision block smooth.Average and variance for the smooth image piece,, the pixel value of these pieces is condensed together, do statistical average, obtain a sample point again according to the similarity of average.Simultaneously, according to the smooth degree of those image blocks that obtain this sample point, calculate the confidence level of this sample point.Like this, the image block that all are smooth all calculates one time, has just obtained the sample value of one group of Noise Estimation, and the confidence level of these sample values.
Then, when reconstruct NLF curve, the different big or small noise level curves to dissimilar CCD carry out principal component analysis (PCA) earlier, obtain the proper vector of these curves, get the dictionary that its topmost 10 proper vectors constitute rarefaction representation.Be used as sampled point to the noise level sample point that has estimated then, utilize the second order cone planing method to obtain treating the rarefaction representation coefficient of estimation curve under this dictionary, obtain the noise level function curve of entire image then with dictionary and multiplication.
Concrete performing step of the present invention is following:
(1) smooth degree of calculating all images piece
The CCD camera is in imaging process; Because photodiode is a device of only supporting single color; Can not distinguish the light of different wave length, need a colour filter Matrix C FA (Color Filter Array), so that let inductor distinguish three kinds of basic colors of red, green, blue of forming visible light.CCD can only catch 25% ruddiness, 50% green glow and 25% blue light like this, and remainder then is to obtain through certain algorithm interpolation.Because the demosaic algorithm can change the smooth degree of image block, particularly for R passage and B passage, influences bigger.Therefore, in the present invention, we only estimate the smooth degree of G passage.When a piece of G passage was considered to smooth, we thought that the R of same position and B channel image piece also are smooth.So both improve smooth search precision, reduced computation complexity again.
What the present invention adopted is a kind of image dissector of structure-oriented, judges whether image block is smooth.What this analyzer adopted is eight high pass operators, comprises along four rectilinear directions of piece and the particular orientation at four angles.The template of 5x5 is as shown in Figure 2.
If the block size that we choose is W=5, then the coordinate when the piece center pixel is that (i, in the time of j), current image block is expressed as
Wherein, (1)
The superscript here (G), expression be the G passage of chromatic noise image.
And the template representation of the smooth analyzer of image is M
n, n=1,2......7,8, then the smooth degree of current block can be weighed and do
And the gray-scale value level of current block, we represent with the average of piece
We are with N
sStep-length in view picture figure, get piece successively, N
sSize depend on and need CCD noise image size, when guaranteeing abundant sampled point, reduce calculated amount as far as possible.
(2) select appropriate threshold, seek all smooth
We need be provided with certain threshold value and judge that which piece is smooth after the complete width of cloth image of traversal.Because the CCD noise depends critically upon signal intensity, if when simply only getting a value, certainly will cause bigger error meeting as threshold value, cause little noise region to cross and estimate to owe estimation with big noise region.Therefore, we adopt a kind of adaptive method to judge here.
Through (1), we have obtained the smooth degree of each piece in the G passage and the grey level of this piece.We will belong to the smooth degree of all pieces of same grey level and do on average, obtain the smooth degree curve of a noise image in whole gray space.The S here
IjThe index set of the similar image block of expression grey level, ε representes the gray-scale deviation that allows, | S
Ij| the expression S set
IjThe number of middle element;
Wherein,
Like this; We are according to the size of current block; Select a suitable parameters λ, then obtain judging the threshold value that image block
is whether smooth
So, we have also just obtained comprising the set of all smooth image blocks
(3) utilize the smooth image piece, the calculating noise sample point
Through (2), we have obtained smooth in R, G, three passages of B.We need pass through these image blocks then, the sample point that comes calculating noise to estimate.Because the influence of demosaic, we can not directly calculate smooth variance as noise variance.Here, we are the arrangement sample modes according to the colour filter Matrix C FA of CCD camera, extract in each passage the pixel without interpolation respectively out, and the gray-scale value level according to pixel place image block constitutes a new set.The sample mode of three passages of note CFA is M
(R), M
(G), M
(B), then the new set that obtains to gray-scale value L does
Wherein,
And
The gray-scale deviation of the permission that the ε here representes, promptly we think that the average gray of smooth image piece drops on [L-ε, L+ ε] when interior, these pieces belong to same grey level.
(4) calculate the confidence level of each sample point
The confidence level of each sample is relevant with the smooth degree of those image blocks that calculate this sample.Image block is smooth more, and the sample point confidence level that obtains is high more.Because the smooth degree of R passage and B passage is to weigh through the G channel image piece of same position.So for each channel image piece on the same position, we give its confidence level that equates, promptly
The rate of decay of the h control confidence level here.
Be similar to the process of sample estimates point in (3), we also can obtain the collection of pixels C of gray-scale value L
LCorresponding confidence level set
Wherein,
(5) calculate the base that is used for rarefaction representation
Through (1)-(4) step, we have obtained the confidence level of Noise Estimation sample point and sample point, but because sample point is not continuous, we need train one group of base for curve Reconstruction.Here, with experimental simulation CCD noise, estimate the noise level function (NLF) of various types of cameras under different noise intensities.The CCD noise model is:
I=f(L+n
s+n
c)+n
q (14)
Wherein, I representes the noise image that actual observation is arrived, the intensity of illumination that the L presentation video is desirable, n
sExpression depends on the noise of intensity of illumination L, n
cExpression does not rely on the noise with intensity of illumination L, n
qThe expression quantizing noise because value is smaller, can be ignored it.And the synthetic process of CCD noise of f () expression, as shown in Figure 3, the CRF here representes camera response function (Camera Response Function).
The CRF function is to obtain from the technical information or the experiment measuring of each camera.The present invention selects preceding 190 functions in the CRF function library for use, and this has comprised the response function of all camera type in the real world basically.And experiment measuring is found, works as σ
s=0.16, σ
c=0.06 o'clock, the noise of image was very big, therefore with the maximal value of these two values as these two parameters.So just obtained each parameter in the CCD noise model.When the input one width of cloth from 0 to 255 gradually change standard picture (like Fig. 4) time, we just can obtain a width of cloth CCD noise image.Utilize following formula like this:
Just can obtain a NLF curve, work as f, n
s, n
cWhen changing successively, just obtained a NLF function library, then its principal component analysis (PCA) has been obtained the proper vector of these curves, got maximum preceding 10 proper vectors of institute's character pair value and constitute dictionary D.
(6) utilize rarefaction representation reconstruct curve
Through (1)-(4) step, we have obtained the confidence level of Noise Estimation sample point and sample point, pass through (5) again, the base of the Noise Estimation of having got back.
Sample point is designated as y ∈ R
M * 1(m<<256), original noise level signal is designated as x ∈ R
256 * 1, the relational expression below x and y satisfy
y=Фx (16)
Ф ∈ R wherein
M * 256, every row has only one 1 element, place row corresponding the gray level at y place (gray level that we adopt is 0 to 255, when drawing the noise level curve, it has been normalized to 0 to 1 interval).Ф is equivalent to a sampling matrix.
The rarefaction representation theory shows, if signal x has rarefaction representation under one group of base D, the reconstruct of signal x just can be converted into the sparse solution of finding the solution the consistent equation of observation so, promptly
s.t.||W·(y-ΦDα)||
2<ε (17)
Wherein α waits to ask the rarefaction representation coefficient of signal x under dictionary D, and W representes the confidence level of each sample, and ε is a very little number, the curve of expression and the approximation ratio of sample.Utilize second order cone planning (second-order cone programming) method to find the solution (17) formula, the rarefaction representation
of noise level function curve x under D that just can obtain CCD then reconstruct CCD noise level function curve is
In the present invention, we have chosen four groups of test pattern goldhill, boats, pen and pepper image, add the noise of dissimilar and different sizes, successively like Fig. 5, Fig. 8, Figure 11 and shown in Figure 14.We utilize picture structure analyzer (like Fig. 2) earlier, seek the smooth image piece, and the variance of obtaining piece then is as noise; The average of these pieces; According to the grey level of piece take statistics average after, obtain the Noise Estimation sample point of these images, like Fig. 6, Fig. 9, Figure 12 and shown in Figure 15.And then utilize the base that extracts from all NLF functions, in conjunction with the rarefaction representation technology, estimate the noise level function of CCD, like Fig. 7, Figure 10, Figure 13 and shown in Figure 16.We can see from figure, the new method that we propose to the CCD noise can estimate very accurate.
Claims (6)
1. the signal dependent noise method of estimation of a CCD image sensor; CCD is the prefix abbreviation of beam coupling device Charge-Coupled Device; It is characterized in that, comprise the steps: to adopt based on the picture structure analyzer of high pass operator and seek smooth of image; Then these pieces are sampled according to the arrangement mode of color filter Matrix C FA; CFA is the prefix abbreviation of Color Filter Array; Obtain a corresponding set of each gray-scale value, calculate the average and the variance of this set, obtain one group of Noise Estimation sample point; According to the smooth degree that obtains that set of diagrams picture piece of each sample point, calculate the confidence level of this sample point simultaneously; The horizontal function NLF of calculating noise function library is done principal component analysis (PCA) to it, the proper vector set that calculates, as the dictionary of rarefaction representation; Combine sample point, sample point confidence level at last and utilize principal component analysis (PCA) in advance, the technology through rarefaction representation reconstructs the noise level curve.
2. the method for claim 1 is characterized in that, adopts to seek smooth of image based on the picture structure analyzer of high pass operator and further be refined as:
11) the picture structure analyzer is made up of eight high pass operators, and image is smooth more, and the value of calculating is more little;
12) in the G passage, get image block
and the step-length of getting piece; Be to follow the big or small W of image block to interrelate; The number that should guarantee piece can not reduce calculated amount very little again as far as possible;
Wherein, (1)
Here
Centre coordinate is positioned at (i, image block j), I in the expression G passage
(G)(k, l) coordinate is (k, pixel value l), W in the expression G passage
IjThe presentation video piece
The coordinate range of pixel;
13) according to the template M of picture structure analyzer
n, n=1,2......7,8, computing center's coordinate be positioned at (i, the smooth degree of image block j), here
The image block of expression and the template of structure analyzer are made convolution;
14) average of computing block is as the grey level of this image block;
According to the smooth degree of piece, calculate and to define the whether smooth threshold curve of image block, find out all smooth image set of blocks in three passages of entire image;
21) to all images piece of G passage,, carry out statistical average, calculate the smooth degree curve of image on whole gray space according to the grey level of piece;
Wherein,
Here, S
IjThe index set of the similar image block of expression grey level, the grey level of L presentation video piece, ε represent the gray-scale deviation that allows, | S
Ij| the expression S set
IjThe number of middle element;
22) according to the size of current block, select a suitable parameters λ, just obtained the judgement image block
Whether smooth threshold tau
Ij
The corresponding smooth degree of grey level
in
expression entire image is here calculated by formula (4);
23) according to threshold curve; Judge whether each G channel image piece is smooth, for being considered to smooth image block, the R passage of same position and the image block in the B passage; Also be considered to smooth, so just obtained the set of all smooth image pieces of entire image;
3. the method for claim 1 is characterized in that, samples according to the arrangement mode of color filter Matrix C FA, obtains a corresponding set of each gray-scale value, calculates the average and the variance of this set, obtains one group of Noise Estimation sample point, is specially:
31) according to the arrangement mode of the colour filter Matrix C FA of CCD camera, to extract out respectively in each passage without the pixel of interpolation, the gray-scale value level according to pixel place image block constitutes a new set, remembers that the sample mode of three passages of CFA is M
(R), M
(G), M
(B), then the new set that obtains to gray-scale value L is:
Wherein,
And
The e here representes the sample mode dot product of each channel image piece and CFA, samples out without the pixel of interpolation;
32) calculate the average and the variance of this set, just obtained the sample point
of a Noise Estimation
The I here
LThe pixel level that expression is calculated through the corresponding collection of pixels of grey level L,
Remarked pixel water
Flat I
LCorresponding noise level.
4. the method for claim 1 is characterized in that, simultaneously according to the smooth degree that obtains that set of diagrams picture piece of each sample point, calculates the confidence level of this sample point, is specially:
41) same images of positions piece in each passage is endowed the confidence level that it equates, promptly
The ω here
IjExpression is centered close to (i, the confidence level of image block j), the rate of decay of h control confidence level;
42) obtain the collection of pixels C of gray-scale value L
LCorresponding confidence level set
Wherein,
43) this confidence level set is asked average, just obtained the corresponding confidence level of this sample point;
5. the method for claim 1 is characterized in that, the horizontal function NLF of calculating noise function library is done principal component analysis (PCA) to it, and the proper vector set that calculates, the dictionary as rarefaction representation is specially:
51) in order to estimate the noise level function NLF of various types of cameras under different noise intensities, the CCD noise model of structure is:
I=f(L
I+n
s+n
c)+n
q (14)
Wherein, I representes the noise image that actual observation is arrived, L
IThe intensity of illumination that presentation video is desirable, n
sExpression depends on intensity of illumination L
INoise, n
cExpression does not rely on and intensity of illumination L
INoise, n
qExpression quantizing noise, and f () expression camera response function CRF;
52) when the input one width of cloth from 0 to 255 gradually change standard picture the time, obtain a width of cloth CCD noise image, utilize following formula, just can obtain a NLF curve;
The f here representes camera response function, σ
sExpression depends on intensity of illumination L
IThe standard deviation of noise, σ
cExpression does not rely on intensity of illumination L
IThe standard deviation of noise, I
N(g) building-up process of expression noise, E [g] expression is asked expectation to the operation result of the inside;
53) according to the situation of the CCD camera of real world, choose the suitable parameters scope, work as f, n
s, n
cWhen changing successively, just obtained a NLF function library;
54) this NLF function library is done principal component analysis (PCA), obtain the eigenwert and its characteristic of correspondence vector of these curves, the dictionary D of the set of proper vector as rarefaction representation.
6. the method for claim 1 is characterized in that, combines sample point, sample point confidence level at last and utilizes principal component analysis (PCA) in advance, and the technology through rarefaction representation reconstructs the noise level curve:
61) sample point is designated as y ∈ R
M * 1, m<<256, the R here
M * 1What represent is the set of m on the real number space * 1 dimensional vector, and original noise level signal is designated as x ∈ R
256 * 1, the relational expression below x and y satisfy
y=Фx (16)
Ф ∈ R wherein
M * 256, every row has only one 1 element, place row corresponding the gray level at y place, be equivalent to a sampling matrix;
62) the rarefaction representation theory shows, if signal x has rarefaction representation under one group of base D, the reconstruct of signal x just can be converted into the sparse solution of finding the solution the consistent equation of observation so, promptly
s.t.||W(y-ФDα)||
2<ε (17)
Wherein, || g||
1With || g||
2Represent L respectively
1Norm and L
2Norm, s.t. representes to make satisfied, and α waits to ask the rarefaction representation coefficient of signal x under dictionary D, and W representes the confidence level of each sample, and ε is a very little number, the curve of expression and the approximation ratio of sample.Utilizing second order cone planning is that second-order cone programming method is found the solution (17) formula, just can obtain the rarefaction representation of noise level function curve x under D of CCD
Then the CCD noise level function curve of reconstruct does
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201210014261.XA CN102609914B (en) | 2012-01-17 | 2012-01-17 | Signal-correlated noise estimating method for image sensor |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201210014261.XA CN102609914B (en) | 2012-01-17 | 2012-01-17 | Signal-correlated noise estimating method for image sensor |
Publications (2)
Publication Number | Publication Date |
---|---|
CN102609914A true CN102609914A (en) | 2012-07-25 |
CN102609914B CN102609914B (en) | 2014-05-28 |
Family
ID=46527260
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201210014261.XA Active CN102609914B (en) | 2012-01-17 | 2012-01-17 | Signal-correlated noise estimating method for image sensor |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN102609914B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103886553A (en) * | 2014-03-10 | 2014-06-25 | 广东威创视讯科技股份有限公司 | Method and system for non-local average value denoising of image |
CN104182948A (en) * | 2013-12-23 | 2014-12-03 | 上海联影医疗科技有限公司 | Estimation method of correlation noise |
CN104581100A (en) * | 2015-02-12 | 2015-04-29 | 张李静 | Color filter array and image processing method |
CN107295217A (en) * | 2017-06-30 | 2017-10-24 | 中原智慧城市设计研究院有限公司 | A kind of video noise estimation method based on principal component analysis |
CN107993200A (en) * | 2017-11-02 | 2018-05-04 | 天津大学 | Picture noise level estimation method based on deep learning |
CN110307804A (en) * | 2019-07-04 | 2019-10-08 | 江南大学 | A kind of curve/Surface quality quantitative evaluation method |
CN110730280A (en) * | 2018-07-17 | 2020-01-24 | 瑞昱半导体股份有限公司 | Noise equalization method and noise removal method |
CN114529509A (en) * | 2022-01-11 | 2022-05-24 | 北京的卢深视科技有限公司 | Image noise evaluation method, electronic device, and computer-readable storage medium |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1867042A (en) * | 2005-05-16 | 2006-11-22 | 安捷伦科技有限公司 | System and method for subtracting dark noise from an image using an estimated dark noise scale factor |
-
2012
- 2012-01-17 CN CN201210014261.XA patent/CN102609914B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1867042A (en) * | 2005-05-16 | 2006-11-22 | 安捷伦科技有限公司 | System and method for subtracting dark noise from an image using an estimated dark noise scale factor |
Non-Patent Citations (5)
Title |
---|
AISHY AMER ET AL: "Fast and Reliable Structure-Oriented Video Noise Estimation", 《IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY》 * |
GLENN E. HEALEY ET AL: "Radiometric CCD Camera Calibration and Noise Estimation", 《IEEE TRANSACTIONS ON PALTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
KENJI IRIE ET AL: "A Technique for Evaluation of CCD Video-Camera Noise", 《IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY》 * |
MIKHAIL L. USS ET AL: "Local Signal-Dependent Noise Variance Estimation From Hyperspectral Textural Images", 《IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING》 * |
MOHAMMED GHAZAL ET AL: "Homogeneity Localization Using Particle Filters With Application to Noise Estimation", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104182948A (en) * | 2013-12-23 | 2014-12-03 | 上海联影医疗科技有限公司 | Estimation method of correlation noise |
CN104182948B (en) * | 2013-12-23 | 2015-07-22 | 上海联影医疗科技有限公司 | Estimation method of correlation noise |
CN103886553A (en) * | 2014-03-10 | 2014-06-25 | 广东威创视讯科技股份有限公司 | Method and system for non-local average value denoising of image |
CN103886553B (en) * | 2014-03-10 | 2017-02-01 | 广东威创视讯科技股份有限公司 | Method and system for non-local average value denoising of image |
CN104581100A (en) * | 2015-02-12 | 2015-04-29 | 张李静 | Color filter array and image processing method |
CN107295217B (en) * | 2017-06-30 | 2020-06-12 | 中原智慧城市设计研究院有限公司 | Video noise estimation method based on principal component analysis |
CN107295217A (en) * | 2017-06-30 | 2017-10-24 | 中原智慧城市设计研究院有限公司 | A kind of video noise estimation method based on principal component analysis |
CN107993200A (en) * | 2017-11-02 | 2018-05-04 | 天津大学 | Picture noise level estimation method based on deep learning |
CN110730280A (en) * | 2018-07-17 | 2020-01-24 | 瑞昱半导体股份有限公司 | Noise equalization method and noise removal method |
CN110730280B (en) * | 2018-07-17 | 2021-08-31 | 瑞昱半导体股份有限公司 | Noise equalization method and noise removal method |
CN110307804A (en) * | 2019-07-04 | 2019-10-08 | 江南大学 | A kind of curve/Surface quality quantitative evaluation method |
CN110307804B (en) * | 2019-07-04 | 2021-03-30 | 江南大学 | Quantitative evaluation method for curve/curved surface quality |
CN114529509A (en) * | 2022-01-11 | 2022-05-24 | 北京的卢深视科技有限公司 | Image noise evaluation method, electronic device, and computer-readable storage medium |
CN114529509B (en) * | 2022-01-11 | 2022-12-16 | 合肥的卢深视科技有限公司 | Image noise evaluation method, electronic device, and computer-readable storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN102609914B (en) | 2014-05-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN102609914B (en) | Signal-correlated noise estimating method for image sensor | |
CN110992275B (en) | Refined single image rain removing method based on generation of countermeasure network | |
Dudhane et al. | RYF-Net: Deep fusion network for single image haze removal | |
CN111709902A (en) | Infrared and visible light image fusion method based on self-attention mechanism | |
CN107680116B (en) | Method for monitoring moving target in video image | |
CN110136060B (en) | Image super-resolution reconstruction method based on shallow dense connection network | |
CN106447668B (en) | The small target detecting method restored under IR Scene based on grab sample and sparse matrix | |
CN107491793B (en) | Polarized SAR image classification method based on sparse scattering complete convolution | |
CN107292830A (en) | Low-light (level) image enhaucament and evaluation method | |
CN107609573A (en) | High spectrum image time varying characteristic extracting method based on low-rank decomposition and empty spectrum constraint | |
CN104517126A (en) | Air quality assessment method based on image analysis | |
CN109345563A (en) | The moving target detecting method decomposed based on low-rank sparse | |
CN105405138A (en) | Water surface target tracking method based on saliency detection | |
CN102436646A (en) | Compressed sensing based CCD (Charge Coupled Device) noise estimation method | |
Wang et al. | Deep near infrared colorization with semantic segmentation and transfer learning | |
CN109615576B (en) | Single-frame image super-resolution reconstruction method based on cascade regression basis learning | |
CN102222321A (en) | Blind reconstruction method for video sequence | |
CN111539434B (en) | Infrared weak and small target detection method based on similarity | |
CN103632373B (en) | A kind of flco detection method of three-frame difference high-order statistic combination OTSU algorithms | |
Azarang et al. | An adaptive multispectral image fusion using particle swarm optimization | |
CN116051444A (en) | Effective infrared and visible light image self-adaptive fusion method | |
Yang et al. | Estimation of signal-dependent sensor noise via sparse representation of noise level functions | |
CN108596831A (en) | A kind of super resolution ratio reconstruction method returned based on AdaBoost examples | |
Zhang et al. | Multisensor Infrared and Visible Image Fusion via Double Joint Edge Preservation Filter and Nonglobally Saliency Gradient Operator | |
Hernandez et al. | Classification of color textures with random field models and neural networks |
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 |