Summary of the invention
The object of the invention is to overcome the deficiency of above-mentioned prior art, proposed a kind of non local image de-noising method based on HWD conversion, to realize the taking into account of edge and smooth region in natural image denoising, improve image denoising effect.
For achieving the above object, the present invention proposes a kind of non local image de-noising method based on HWD conversion, comprises the steps:
(1) using the image block in the noisy image of input as with reference to piece, for this reference block, according to Euclidean distance formula, calculate the distance of all image blocks in this reference block and its neighborhood, get distance and be less than threshold value
image block form similar group of this reference block
Wherein, X is noisy image, and x is the pixel in X,
for reference block,
be with
centered by neighborhood in image block,
for the size of image block,
to judge two threshold values whether image block is similar;
(2) image block in similar group is carried out to two-dimensional wavelet transformation, and the large coefficient in wavelet coefficient and little coefficient are labeled as respectively to 1 and 0, according to 0 and 1 of each subband on the thickest yardstick of all image blocks in similar group distribution, judge the singularity of similar group;
(3) according to the singularity of similar group, simple similar group of singularity is carried out to two-dimensional wavelet transformation and interblock one-dimensional wavelet transform, make TIHWD conversion and interblock one-dimensional wavelet transform for similar group of singularity complexity, hard-threshold contracted transformation coefficient also carries out corresponding inverse transformation, obtains the estimated value of image block in similar group
Wherein,
represent the three-dimension varying being formed by two-dimensional transform and interblock one-dimensional wavelet transform,
for inverse transformation, Υ represents hard-threshold shrinkage operation, and general value is λ σ, and λ is the artificial constant of setting, and σ is that noise criteria is poor;
(4) according to the singularity of similar group, the estimated value of integral image piece, obtains basic estimated image
Wherein,
it is similar group
corresponding weights,
with pixel x
mfor the image block on left summit
fundamental function, pixel
time,
value is 1, otherwise is 0;
Wherein,
it is similar group
hard-threshold is shunk the number of rear nonzero coefficient;
(5) at basic estimated image
upper, build its similar group by block matching method
similar group of manual record
in the coordinate information of similar:
Wherein,
with
it is basic estimated image
in reference block and candidate blocks,
the size of image block,
to judge two threshold values whether image block is similar;
(6) according to similar group
in the coordinate information of similar, from noisy image, extract the image block of answering in contrast, obtain similar group of noisy image
to basic estimated image
similar group
with similar group of noisy image
all carry out three-dimension varying, according to
conversion coefficient calculate Wei Na shrink matrix W, will
conversion coefficient and after W pointwise multiplies each other, then carry out inverse transformation, obtain similar group
the estimated value of middle image block
Wherein,
represent the three-dimension varying being formed by two-dimension discrete cosine transform DCT and interblock one-dimensional wavelet transform,
represent inverse transformation, W is that Wei Na shrinks matrix;
(7) estimated value to the image block obtaining
weighted mean, obtains final denoising image
Wherein,
it is similar group
corresponding weights,
with pixel x
mfor the image block on left summit
fundamental function, pixel
time,
value is 1, otherwise is 0;
Wherein, W is similar group
corresponding Wei Na shrinks matrix.
In above-mentioned steps (2), judge the singularity of similar group, obtain in the following manner:
(1) in similar group similar carry out three layers of wavelet decomposition, by the large coefficient mark in the wavelet coefficient after conversion
Be 1, the little coefficient of large coefficient in the wavelet coefficient after conversion be labeled as to 0:
Wherein, ω is wavelet coefficient,
be the mark value of wavelet coefficient, thresh is threshold value, and value is
(2) distribute according to 0 and 1 of each subband on all similar yardsticks the thickest in similar group, judge the singularity of similar group:
If all image blocks meet in similar group: the low frequency on the thickest yardstick, the mark value of vertical and horizontal high-frequency sub-band coefficient are 1, this similar group is judged as similar group of singularity complexity; If do not meet above-mentioned requirements, this similar group is judged as simple similar group of singularity.
The present invention has the following advantages compared with prior art:
1. the present invention carries out mark by the wavelet coefficient to similar, can judge exactly the singularity of similar group, and then different similar group selections is applicable to self conversion to singularity.
2. the present invention is according to the singularity of similar group, and to the simple similar group selection wavelet transformation of singularity, the similar group selection TIHWD conversion of singularity complexity, combines two kinds of conversion advantage separately, when denoising, and keep the edge information, grain details better.
The present invention due to similar group selection adapt to the conversion of its singularity features, can obtain basic estimated image more accurately, and then improve and in subordinate phase, build the accuracy that similar group and Wei Na shrink matrix, thereby improve the final denoising effect of image.
Accompanying drawing explanation
Fig. 1 is process flow diagram of the present invention;
Fig. 2 is the test pattern that the present invention tests use;
Fig. 3 is three layers of wavelet decomposition figure of image block of 8 × 8;
Fig. 4 is σ=25 o'clock, denoising result by existing three kinds of denoising methods to Lena image, wherein (a) figure is original Lena figure, (b) be noisy Lena figure, (c) be TIHWD bivariate threshold method denoising result, (d) being the denoising result of BM3D method, is (e) denoising result of the present invention;
Fig. 5 is σ=50 o'clock, denoising result by existing three kinds of denoising methods to house image, wherein (a) figure is original house figure, (b) be noisy house figure, (c) be TIHWD bivariate threshold method denoising result, (d) being the denoising result of BM3D method, is (e) denoising result of the present invention;
Fig. 6 is σ=35 o'clock, denoising result by existing three kinds of denoising methods to Baboon image, wherein (a) figure is original Baboon figure, (b) be noisy Baboon figure, (c) be TIHWD bivariate threshold method denoising result, (d) being the denoising result of BM3D method, is (e) denoising result of the present invention;
Fig. 7 is σ=50 o'clock, denoising result by existing three kinds of denoising methods to fingerprint image, wherein (a) figure is original fingerprint figure, (b) be noisy fingerprint figure, (c) be TIHWD bivariate threshold method denoising result, (d) being the denoising result of BM3D method, is (e) denoising result of the present invention;
Fig. 8 is partial enlarged drawing corresponding to denoising result figure in Fig. 5;
Fig. 9 is the corresponding partial enlarged drawing of denoising result figure in Fig. 7.
Embodiment
HWD(HybridWaveletsanddirectionalfilterbanks) be a kind of rarefaction representation method of break-even characteristics of image, it can extract profile and the detailed information of image effectively.
With reference to accompanying drawing 1, the invention provides a kind of non local image de-noising method based on HWD conversion, comprise the steps:
Step 1, adds to the test pattern in Fig. 2 the noise that standard deviation is σ, obtains noisy image X, builds its similar group:
1.1) in noisy image X, with step-length
get reference block
x
rfor
left summit, with x
rfor in
In 39 × 39 neighborhood of the heart, according to the distance of candidate blocks in Euclidean distance formula computing reference piece and this neighborhood
Wherein,
for reference block,
for candidate blocks,
the size of image block, σ≤40 o'clock,
value is 8,
be 3; When σ > 40,
value is 12,
be 4;
1.2) get distance
d is less than threshold value
similar group of candidate blocks composition reference block:
Wherein,
to judge two threshold values whether image block is similar, σ≤40 o'clock,
value is 2500; When σ > 40,
value is 5000;
Step 2, carries out two-dimensional wavelet transformation to image block in similar group, according to the large coefficient in wavelet coefficient and the distribution of little coefficient, judges the singularity of similar group:
2.1) image block in similar group is carried out to three layers of bior1.5 wavelet decomposition, the large coefficient in wavelet coefficient is labeled as to 1, the little coefficient in wavelet coefficient is labeled as to 0:
Wherein, ω is wavelet coefficient,
be the wavelet coefficient after mark, thresh is threshold value, and value is
2.2) according to 0 and 1 distribution mark value of low frequency, vertical direction and horizontal direction high-frequency sub-band on the thickest yardstick of all image blocks in similar group, judge the singularity of similar group:
As Fig. 3, if all image blocks meet in similar group: the low frequency LL on the thickest yardstick
3, vertical direction LH
3with horizontal direction HL
3the mark value of high-frequency sub-band coefficient is 1, and this similar group is judged as similar group of singularity complexity; If do not meet above-mentioned requirements, this similar group is judged as simple similar group of singularity.
Step 3, carry out the collaborative filtering of transform domain to similar group:
3.1) carry out the TIHWD conversion HWD of the translation invariant form (conversion) and interblock one-dimensional wavelet transform to similar group of singularity complexity, simple similar group of singularity is carried out two-dimensional wavelet transformation (or dct transform) and interblock one-dimensional wavelet transform;
3.2) hard-threshold is shunk three-dimension varying coefficient, remakes corresponding contrary three-dimension varying:
Wherein,
represent the three-dimension varying being made up of two-dimensional transform (small echo, TIHWD or two-dimension discrete cosine transform DCT) and interblock one-dimensional wavelet transform, Υ is hard-threshold shrinkage operation, λ
3Dσ is threshold value, similar group of singularity complexity: λ
3Dvalue is 3 (σ≤40 o'clock) and 3.1 (when σ > 40), simple similar group of singularity: λ
3Dvalue is 2.7 (when σ <40) and 2.8 (when σ > 40), and σ is that noise criteria is poor,
for inverse transformation;
Step 4, according to the singularity of similar group, the estimated value of integral image piece, obtains basic estimated image:
Add up respectively each pixel, the number of the similar group of estimated value of returning that singularity is different, gets estimated value weighted mean corresponding to more person, as the estimated value of pixel, obtains basic estimated image
Wherein,
it is similar group
corresponding weights,
with pixel x
mfor the image block on left summit
fundamental function, pixel
time,
value is 1, otherwise is 0;
Wherein,
it is similar group
hard-threshold is shunk the number of rear nonzero coefficient;
Step 5, on basic estimated image, builds its similar group by block matching method
the coordinate information of similar in similar group of manual record:
Wherein,
with
be reference block and the candidate blocks in basic drawing for estimate, the size of piece is
to judge two threshold values whether image block is similar, and the step-length here
value is 6, σ≤40 o'clock,
value is 400,
value is 8; When σ > 40,
value is 3500,
value is 11;
Step 6, the coordinate information obtaining according to step 5 extracts corresponding image block from noisy image, obtains similar group
carry out the collaborative filtering of transform domain:
6.1) similar group to basic estimated image
carry out three-dimension varying, calculate Wei Na according to conversion coefficient and shrink matrix:
Wherein,
represent the three-dimension varying being formed by two-dimension discrete cosine transform DCT and interblock one-dimensional wavelet transform;
6.2) to similar group
carry out same three-dimension varying, will
conversion coefficient and after W pointwise multiplies each other, then inverse transformation, obtains similar group
the estimated value of middle image block:
Wherein, W is that Wei Na shrinks matrix,
represent contrary three-dimension varying;
Step 7, to similar group
middle image block estimated value weighted mean, obtains final denoising image
Wherein,
it is similar group
corresponding weights,
with pixel x
mfor the image block on left summit
fundamental function, pixel
time,
value is 1, otherwise is 0;
Wherein, W is similar group
corresponding Wei Na shrinks matrix.
Effect of the present invention can further confirm by following experiment:
One. experiment condition and content
Experiment condition: as shown in Figure 2, in experiment, denoising method is all to use matlab Programming with Pascal Language to realize to the test pattern that experiment is used.
Experiment content: under above-mentioned experiment condition, 1. bivariate threshold value, BM3D denoising method and the present invention test to use respectively TIHWD.Wherein, the decomposition number of plies in TIHWD bivariate threshold method is [33000], and what BM3D method was used is the parameter in its corresponding article, in the first stage of the present invention, σ≤40 o'clock, two-dimensional transform is bior1.5 wavelet transformation and TIHWD conversion, and when σ > 40, two-dimensional transform is dct transform and TIHWD conversion, the decomposition number of plies of TIHWD conversion is [2,0], one-dimensional wavelet transform is haar wavelet transformation, and the parameter of subordinate phase all arranges consistent with parameter in BM3D.
Two. experimental result
To the denoising result of the test pattern in Fig. 2 as shown in FIG. 4,5,6, 7, wherein Fig. 4 (a) is test pattern Lena, size is that 256 × 256, Fig. 4 (b) is that Fig. 4 (a) is added to standard deviation is the noisy Lena image of 25 noises; Fig. 5 (a) is test pattern house, and size is that 256 × 256, Fig. 5 (b) is that 5 (a) are added to standard deviation is the noisy house image of 50 noises; Fig. 6 (a) is test pattern Baboon, and size is that 512 × 512, Fig. 6 (b) is that 6 (a) are added to standard deviation is the noisy Baboon image of 35 noises; Fig. 7 (a) is test pattern fingerprint, and size is that 512 × 512, Fig. 7 (b) is that 7 (a) are added to standard deviation is the noisy fingerprint image of 50 noises; At Fig. 4 in Fig. 7, (c) be the denoising image that TIHWD bivariate threshold method obtains, (d) be the denoising image that BM3D method obtains, (e) be the denoising image that the present invention obtains, Fig. 8 and Fig. 9 are respectively the partial enlarged drawing of corresponding denoising result in Fig. 5 and Fig. 7.
Use TIHWD bivariate threshold method denoising result in Fig. 7, as shown in (c), can find out from result figure and partial enlarged drawing as Fig. 4, the method can keep detailed information preferably, but in image, has produced obvious Gibbs' effect;
Use BM3D method denoising result if Fig. 4 is in Fig. 7 as shown in (d), can find out from result figure and partial enlarged drawing, the method all has denoising performance preferably at homogeneous area and fringe region, compare TIHWD bivariate threshold method, in result figure, almost there is no Gibbs' effect, but in denoising, can make fuzzy and the continuity that can not keep the edge information of part edge;
The inventive method denoising result is if Fig. 4 is in Fig. 7 as shown in (e), can find out from result figure and partial enlarged drawing, than two kinds of above-mentioned denoising methods, the present invention is removing in noise, the detailed information such as keep the edge information and unsmooth region well.
Quantitative evaluation index by the Y-PSNR PSNR value of each figure in Fig. 2 as denoising result, its computing method are:
U (i) is the original nothing figure that makes an uproar,
for the result figure after denoising, u
max=max{u (i), i ∈ N}, N presentation video size.The PSNR value of above-mentioned existing two kinds of denoising methods and denoising result of the present invention is listed in table 1.
The contrast of table 1 denoising result
From testing, the present invention is with respect to TIHWD bivariate threshold method and BM3D denoising method, and the visual effect of denoising figure and PSNR evaluation index all increase, when the present invention can smooth noise, the edge and the grain details that keep better image, have better denoising effect.
In literary composition, there is 1. coming from foreign language document: EslamiR, RadhaH., Anewfamilyofnonredundanttransformsusinghybridwaveletsand directionalfilterbanks.IEEETrans.ImageProcessing, 2007,16 (4): 1152-1167.