Background technology
Open now and popular denoising digital picture algorithm majority is based on conversion filtering, like wavelet transformation, and wavelet-like transform, discrete cosine transform etc.In transform domain, model can scrutinized and set up to the statistical property of conversion coefficient to help the filtering of adjustment threshold values.In present denoising digital picture algorithm, these conversion usually were complete, promptly used basic conversion a plurality of versions that translation forms in spatial domain to carry out filtering, then with synthesizing final noise reduction result after the filtered inverse transformation.Denoising digital picture system based on crossing complete conversion can describe with Fig. 1.In three modules shown in Figure 1, cross complete conversion, a lot of research explorations has been done in statistical study modeling and filtering scholars in recent years, handles but then just simply get average mostly in synthetic this module of pixel.
At IEEE Transations on lmage Proceedings Vol.16.No.12; O.G.Guleryuz has proposed a kind of pixel weighting composition algorithm based on the hard threshold values filtering of crossing complete orthonormal transformation (like discrete cosine transform) among the Dec.2007, and details are as follows for its complete scheme:
Suppose that original noise-free picture is x; Noise image is y=x+w, and wherein w is that variance is the Gaussian noise of
.Concerning pixel n; Passing through complete orthonormal transformation; After hard threshold values filtering and the inverse transformation, obtain M noise reduction estimated value:
wherein the complete orthonormal transformation of mistake comprises the individual basic conversion of M.Final noise reduction result is synthetic by these estimated value weightings:
The pairing weight r of estimated value that each basic conversion obtains
i(n) obtain with following equation: G
nR (n)=C
nu
R (n) is r
i(n) the M*1 rank matrix that constitutes.C
nBe that zoom variables is to guarantee weight r
i(n) and be 1.U is M*1 rank unit matrixs.G
nBe M*M rank matrix, its element does
H wherein
iAnd H
jBe respectively i and j the pairing matrix of coefficients of basic conversion, S
iAnd S
jBe respectively after i and j basic conversion, the hard pairing index variation matrix of threshold values filtering, and δ
nIt is the impulse function of pixel n correspondence in conversion.If G
nOrder be non-zero, then separate above-mentioned equation and can obtain weight r
i(n); If G
nOrder be zero, weight r then
i(n) all be made as 1/M, promptly get average and synthesize.For the conversion of standard quadrature piece, it is following that O.G.Guleryuz has also proposed a kind of reduction procedure:
The weight r of the Filtering Estimation value that pixel n obtains through i basic conversion
i(n) do
C
n/N(b
n,i)
N (b wherein
N, i) be for filtered nonzero coefficient number in i the transform block at pixel n place, C
nBe that zoom variables is to guarantee weight r
i(n) and be 1.
A kind of pixel weighting synthetic method that O.G.Guleryuz proposes based on the hard threshold values filtering of crossing complete orthonormal transformation; Improved and crossed the synthetic method of just simply getting average of pixel in the complete conversion filtering in the past; Its noise reduction reaches best level in the present disclosed denoising digital picture algorithm; But in the denoising digital picture practice, still exist a lot of noises not remove, main cause is that the noise that the digital picture inductor produces not exclusively satisfies Gauss model, after the Flame Image Process through the IMAQ terminal; Noise is often to medium and low frequency section diffusion, and stronger correlativity in the presentation space territory.
Summary of the invention
Exist a lot of noises not have the deficiency of removing in order to overcome existing pixel weighting synthetic method based on the hard threshold values filtering of crossing complete orthonormal transformation; The invention provides a kind of based on multiresolution and digital image noise reduction method and the system of crossing complete conversion; These method and system are further optimized based on the weighted filtering method of crossing complete conversion, realize noise reduction better.
In order to solve its technical matters, the invention provides a kind of denoising digital picture system based on multiresolution and the complete conversion of mistake, this system comprises: the image smoothing module; The image drop sampling module; Image rises sampling module, and rim detection and area dividing module are crossed complete orthonormal transformation module; Statistical study modeling and filtration module, inverse transformation and pixel synthesis module; This system is in multiresolution decomposition and building-up process; Adopt the image smoothing module and fall sampling module noise image is decomposed into the multiresolution pyramid structure, adopt image to rise sampling module and the processing of filtering noise reduction is shone upon back preceding single order image and obtained final noise reduction image by rank; Under the resolution of every rank, adopt different transform block size to carry out complete orthonormal transformation according to the differentiation result of rim detection and area dividing module, after threshold filter and inverse transformation, with the synthetic noise reduction image that obtains each rank of pixel weighting.On the s rank of multiresolution analysis, input picture I
sPass through image smoothing and fall sampling module and obtain s+1 rank input picture I
S+1S+1 rank noise reduction image
Deduct s+1 rank input picture I
S+1, rise then sampling and with s rank input picture I
sAddition promptly is the pending image I of the noise reduction image mapped on s+1 rank to the s+1 rank
sOn; Result after the addition is carried out complete conversion weighted filtering, obtained s rank noise reduction image
When s>0, s rank noise reduction image
Continuing to feed back to the s-1 rank further handles; When s=0, s rank noise reduction image
It promptly is final noise reduction image.
Alternatively, the image smoothing module can adopt implementations such as weighted mean filtering, Gauss's LPF and mean filter.
Alternatively, the image drop sampling module can adopt the image drop sampling implementation that does not produce aliasing effect, such as: carry out LPF earlier to guarantee to satisfy nyquist sampling theorem, the integer value of taking out then.
Alternatively, image rises the adoptable implementation of sampling module and is: bilinear interpolation, consecutive point repeat interpolation, two cubes of interpolation etc.
Alternatively; Crossing the transform block size of complete orthonormal transformation regulates according to the judgement of rim detection and area dividing module; That the complete conversion of the mistake that the pixel that is judged to be edge feature is carried out is used is less (such as; 4 * 4 or 8 * 8) that transform block, the complete conversion of the mistake that the pixel that is judged to be smooth features is carried out use is big (such as, 16 * 16 or bigger) transform block.Rim detection and area dividing module can adopt Canny, Sobel, and methods such as LoG realize rim detection; Cross complete orthonormal transformation module except the discrete cosine orthogonal transformation that can adopt the O.G.Guleryuz proposition, also can adopt other orthonormal transformations; Statistical study modeling and filtration module also can train the hard-threshold or the soft-threshold filtering method of threshold parameter according to real image except adopting the hard-threshold filtering method based on Gaussian noise of O.G.Guleryuz proposition; A kind of in synthetic complete scheme of the pixel weighting that inverse transformation and pixel synthesis module can adopt O.G.Guleryuz to propose or the reduction procedure.
It is a kind of based on multiresolution and the digital image noise reduction method of crossing complete conversion that the present invention also provides, and the concrete steps of this method comprise:
Step 1 adopts image smoothing and image drop sampling that noise image is resolved into the multiresolution pyramid structure, obtains the input picture of each rank resolution, begins execution in step two from high-order;
Step 2 judges that whether current rank are high-order, and as not, then execution in step three; In this way, execution in step four then;
Step 3 is used single order input picture on the noise reduction figure image subtraction of single order, and the image that obtains is risen after the sampling and the input picture addition of current rank through image;
Step 4 adopts Canny with the input picture of high-order resolution or the image that step 3 obtains, Sobel, and edge detection methods such as LoG are carried out edge detection, and mark off edge feature pixel and level and smooth feature pixel;
Step 5; Selected the size of complete transform block according to the area dividing result of step 4: if pixel is judged to be the edge feature point; Used when then carrying out complete conversion less (such as, 4 * 4 or 8 * 8) transform block, if pixel is judged to be the smooth features point; Used when then carrying out complete conversion big (such as, 16 * 16 or bigger) transform block;
Step 6; The complete transform block of selecting according to step 5 of mistake carried out complete conversion to the input picture of high-order resolution or the image that step 4 obtains; This step also can adopt other orthonormal transformations except the discrete cosine orthogonal transformation that can adopt the O.G.Guleryuz proposition;
Step 7; The image that step 6 is obtained carries out statistical study modeling and filtering; This step also can train the hard-threshold or the soft-threshold filtering method of threshold parameter according to real image except adopting the hard-threshold filtering method based on Gaussian noise of O.G.Guleryuz proposition;
Step 8, the image that step 7 is obtained carry out inverse transformation and pixel is synthetic, a kind of in synthetic complete scheme of the pixel weighting that this step can adopt O.G.Guleryuz to propose or the reduction procedure;
Step 9 is stored current rank noise reduction image;
Step 10 judges whether the sequence number on current rank equals 0, and in this way, then execution in step 11; As not, then will descend single order as current rank, return execution in step two;
Step 11, output noise reduction image.
Alternatively, image smoothing can adopt implementation methods such as weighted mean filtering, Gauss's LPF and mean filter.
Alternatively, image rises the method for sampling to be had: bilinear interpolation, consecutive point repeat interpolation, two cubes of interpolation etc.
Compared with prior art; The present invention combines multiresolution analysis with the pixel weighting synthetic method of crossing complete conversion threshold filter; Under a plurality of resolution, can probe into the spatial domain correlativity of noise to image analysis, thus more effective elimination noise, the particularly noise of medium and low frequency section.Fringe region and the smooth region orthonormal transformation with use different masses size is distinguished in the guidance of surveying on the edge of down, thereby under the prerequisite of protection edge feature, further scope is investigated with the elimination noise in the part of expanded view image space information; And the present invention can filtering noise be not limited to Gaussian noise, can train the noise that threshold parameter adopts soft or hard-threshold filtering method filtering respective type according to real image.
Embodiment
The present invention desires under a plurality of resolution, image to be analyzed; Expectation can be probed into the spatial domain correlativity of noise; Thereby more effective elimination noise; So the present invention has adopted the multiresolution analysis method of using always in the Digital Image Processing, can investigate the characteristics of image under the different spaces yardstick to image analysis under a plurality of resolution, thereby enhancement process effect; And fringe region and the smooth region orthonormal transformation with use different masses size is distinguished in the guidance of surveying on the edge of down, thereby under the prerequisite of protection edge feature, further scope is investigated with the elimination noise in the part of expanded view image space information.
Below in conjunction with accompanying drawing the specific embodiment of the invention is described.
Seeing also shown in Figure 2ly, is of the present invention based on multiresolution analysis and the denoising digital picture system schematic 2 of crossing complete conversion weighted filtering, and this system comprises image smoothing module 201; Image drop sampling module 202; Image rises sampling module 203, and rim detection and area dividing module 204 are crossed complete orthonormal transformation module 205; Statistical study modeling and filtration module 206, inverse transformation and pixel synthesis module 207.This system is in multiresolution decomposition and building-up process; Adopt image smoothing module 201 and fall sampling module 202 noise image is decomposed into the multiresolution pyramid structure, adopt image to rise sampling module 203 and the processing of filtering noise reduction is shone upon back preceding single order image and obtained final noise reduction image by rank; Under the resolution of every rank, adopt different transform block size to carry out complete orthonormal transformation according to the differentiation result of rim detection and area dividing module 204, after threshold filter and inverse transformation, with the synthetic noise reduction image that obtains each rank of pixel weighting.
Suppose that original noise image is I
0The input picture I on the s rank of multiresolution analysis
sPass through
image smoothing module 201 and fall
sampling module 202 processing and obtain s+1 rank input picture I
S+1, then the noise reduction image on s+1 rank
Deduct s+1 rank input picture I
S+1The image that obtains carries out liter sampling and and I through rising
sampling module 203
sAddition promptly is mapped to the noise reduction process on s+1 rank the pending image I on s+1 rank
sOn, the image after the addition was sent into complete conversion weighted
filtering module 204 handle and obtain the noise reduction image
The time,
Continuing to feed back to the s-1 rank further handles; During s=0,
It promptly is final noise reduction image.Wherein the exponent number that need carry out of multiresolution analysis is decided according to the statistical property of image size and noise, and the value of the image correspondence of general 1024 * 768 resolution is [3,5].
Image smoothing module 201 can adopt implementations such as weighted mean filtering, Gauss's LPF and mean filter; Wherein the concrete implementation of weighted mean filtering is following: let P
s(x y) represents I
sLast coordinate is (x, pixel value y), its smooth value P in the horizontal direction
S, h(x is to be that the center is the three-point weight mean value of step-length with step with this some y), that is:
P
s,h(x,y)=(w*P
s(x,y)+P
s(x-step,y)+P
s(x+step,y))/(w+2);
Wherein w is a center weight.In like manner, the smooth value P on the vertical direction
S, v(x is y) according to P
S, h(x, y) calculate and get:
P
s,v(x,y)=(w*P
s,h(x,y)+P
s,h(x,y-step)+P
s,h(x,y+step))/(w+2)
P
S, v(x y) is coordinate for (wherein parameter step length step and center weight w are according to the image size for x, level and smooth back pixel value y), and training such as the actual characteristic of picture noise obtain, and general step-length step value is 1, and center weight w value is 2.
Image drop sampling module 202 can adopt the image drop sampling implementation that does not produce aliasing effect, such as: carry out LPF earlier to guarantee to satisfy nyquist sampling theorem, the integer value of taking out then.
Image rises sampling module 203 adoptable implementations: bilinear interpolation, consecutive point repeat interpolation, two cubes of interpolation etc.
Crossing the transform block size of complete orthonormal transformation regulates according to the judgement of rim detection and area dividing module 204; That the complete conversion of the mistake that the pixel that is judged to be edge feature is carried out is used is less (such as; 4 * 4 or 8 * 8) transform block; That the complete conversion of the mistake that the pixel that is judged to be smooth features is carried out is used is big (such as, 16 * 16 or bigger) transform block.Rim detection and area dividing module 204 can adopt Canny, Sobel, and methods such as LoG realize rim detection; Cross complete orthonormal transformation module 205 except the discrete cosine orthogonal transformation that can adopt the O.G.Guleryuz proposition, also can adopt other orthonormal transformations; Statistical study modeling and filtration module 206 also can train the hard-threshold or the soft-threshold filtering method of threshold parameter according to real image except adopting the hard-threshold filtering method based on Gaussian noise of O.G.Guleryuz proposition; A kind of in synthetic complete scheme of the pixel weighting that inverse transformation and pixel synthesis module 207 can adopt O.G.Guleryuz to propose or the reduction procedure.
Seeing also shown in Figure 3ly, is of the present invention based on the process flow diagram 3 of multiresolution analysis with the digital image noise reduction method of crossing complete conversion weighted filtering, and this process flow diagram comprises following concrete steps (is that example describes with the s rank):
Step S301 adopts image smoothing and image drop sampling that noise image is resolved into the multiresolution pyramid structure, obtains the input picture of each rank resolution, begins execution in step S302 from high-order.
Step S302 judges that whether current rank (s rank) are high-order, as not, and execution in step S303 then; In this way, execution in step S304 then.
Step S303 uses single order input picture on the noise reduction figure image subtraction of single order (s+1 rank), and the image that obtains is risen after the sampling and the input picture addition of current rank through image;
The adoptable image of this step rises the method for sampling to be had: bilinear interpolation, and bicubic interpolation, consecutive point repeat, two cubes of differences etc.
Step S304 adopts Canny with the input picture of high-order resolution or the image that step S303 obtains, Sobel, and edge detection methods such as LoG are carried out edge detection, and mark off edge feature pixel and level and smooth feature pixel.
Step S305; Selected the size of complete transform block according to the area dividing result of step S304: if pixel is judged to be the edge feature point; Used when then carrying out complete conversion less (such as, 4 * 4 or 8 * 8) transform block, if pixel is judged to be the smooth features point; Used when then carrying out complete conversion big (such as, 16 * 16 or bigger) transform block.
Step S306; The complete transform block of selecting according to step S305 of mistake carried out complete conversion to the input picture of high-order resolution or the image that step S304 obtains; This step also can adopt other orthonormal transformations except the discrete cosine orthogonal transformation that can adopt the O.G.Guleryuz proposition.
Step S307; The image that step S306 is obtained carries out statistical study modeling and filtering; This step also can train the hard-threshold or the soft-threshold filtering method of threshold parameter according to real image except adopting the hard-threshold filtering method based on Gaussian noise of O.G.Guleryuz proposition.
Step S308, the image that step S307 is obtained carry out inverse transformation and pixel is synthetic, a kind of in synthetic complete scheme of the pixel weighting that this step can adopt O.G.Guleryuz to propose or the reduction procedure.
Step S309 stores current rank noise reduction image.
Step S310 judges whether the sequence number s on current rank equals 0, in this way, and execution in step S311 then; As not, then will descend single order s-1 as current rank, return execution in step S302.
Step S311, output noise reduction image.
In above-mentioned steps S301, to s rank input picture I
sCarry out image smoothing, can adopt implementation methods such as weighted mean filtering, Gauss's LPF and mean filter, the concrete disposal route of its weighted mean filtering is following: let P
s(x y) represents s rank input picture I
sLast coordinate is (x, pixel value y), its smooth value P in the horizontal direction
S, h(x is to be that the center is the three-point weight mean value of step-length with step with this some y), that is:
P
s,h(x,y)=(w*P
s(x,y)+P
s(x-step,y)+P
s(x+step,y))/(w+2);
Wherein w is a center weight; Smooth value P on its vertical direction
S, v(x is y) according to P
S, h(x, y) calculate and get:
P
s,v(x,y)=(w*P
s,h(x,y)+P
s,h(x,y-step)+P
s,h(x,y+step))/(w+2)
P
S, v(x y) is coordinate for (wherein parameter step length step and center weight w are according to the image size for x, level and smooth back pixel value y), and training such as the actual characteristic of picture noise obtain, and general step-length step value is 1, and center weight w value is 2.
And image drop sampling can adopt the image drop sampling implementation that does not produce aliasing effect, such as: carry out LPF earlier to guarantee to satisfy nyquist sampling theorem, the integer value of taking out then.
This method combines multiresolution analysis with the pixel weighting synthetic method of crossing the filtering of complete conversion hard-threshold; In resolution decomposition and building-up process; Adopt image smoothing and fall the method for sampling picture breakdown is become the pyramid structure of multiresolution, adopt to rise the method for sampling and pursue rank the processing of filtering noise reduction is shone upon back preceding single order image and obtained final noise reduction image; Under the resolution of every rank, adopted complete conversion, after hard-threshold filtering and the inverse transformation, with the synthetic noise reduction image that obtains each rank of pixel weighting.When s>0, S rank noise reduction image
continues to feed back to the s-1 rank further to be handled; When s=0, S rank noise reduction image
promptly is final noise reduction image.Wherein the exponent number that need carry out of multiresolution analysis is decided according to the statistical property of image size and noise, and the value of the image correspondence of general 1024 * 768 resolution is [3,5].
Compared with prior art; The present invention combines multiresolution analysis with the pixel weighting synthetic method of crossing complete conversion threshold filter; Under a plurality of resolution, can probe into the spatial domain correlativity of noise to image analysis, thus more effective elimination noise, the particularly noise of medium and low frequency section.Fringe region and the smooth region orthonormal transformation with use different masses size is distinguished in the guidance of surveying on the edge of down, thereby under the prerequisite of protection edge feature, further scope is investigated with the elimination noise in the part of expanded view image space information; And the present invention can filtering noise be not limited to Gaussian noise, can train the noise that threshold parameter adopts soft or hard-threshold filtering method filtering respective type according to real image.
It is understandable that, concerning those of ordinary skills, can be equal to replacement or change according to technical scheme of the present invention and inventive concept thereof, and all these changes or replacement all should belong to the protection domain of the appended claim of the present invention.