CN105894478A - Image denoising method based on statistical local rank characteristics - Google Patents

Abstract

The invention discloses an image denoising method based on statistical local rank characteristics. The image denoising method comprises the following steps: performing local rank transformation on the image under different parameter conditions by utilizing a local rank operator, thereby obtaining positive local rank transformation and negative local rank transformation of the image; adding the local rank transformation and negative local rank transformation to obtain statistical local rank characteristics with continuous parameter changes; taking the statistical local rank characteristics as constraint conditions on the basis of an image denoising method with sparse representation, and performing primary denoising on the image; and finally, performing secondary denoising on the image by controlling the difference of the statistical local rank characteristics between the images before and after denoising, and removing the image noise, thereby obtaining a final clear image. The method has the obvious effects that compared with the traditional denoising method based on sparse representation, the method has a better denoising effect, can acquire a denoised image with high quality, and further can effectively guarantee the reliability of subsequent image processing and analyzing.

Description

Image de-noising method based on statistics local order feature

Technical field

The present invention relates to technical field of image processing, specifically, be a kind of based on statistics local order feature Image de-noising method.

Background technology

Image is obtaining and in transmitting procedure, is inevitably being affected by various noises, cause figure picture element Amount declines, it is impossible to meet the demand of subsequent treatment.In order to improve picture quality, Image Denoising Technology meet the tendency and Raw.

In recent years, the strong tools that the rarefaction representation of signal becomes higher-dimension signal acquisition, characterizes and compress. Sparse representation model assumes that the non-noise composition in image can be by rarefaction representation, and noise contribution can not be dilute Dredging and represent, researcher utilizes this characteristic of sparse representation method to carry out numerous studies.Beams etc. are strange at K- On the basis of different value decomposition algorithm, integrated structure cluster and dictionary learning, devise a kind of based on non local just Then changing the image de-noising method of rarefaction representation, compare traditional K-singular value decomposition method, the method can be more preferably Ground holding image structure information (see document " Shearlet territory SAR image denoising based on rarefaction representation ", It is published in electronics and information journal, the 9th phase in 2012).Grandsons etc. use a kind of shellfish based on sparse representation model This maximum a-posteriori estimation method of leaf is removed picture noise and (is seen document " A novel image denoising Algorithm using linear Bayesian MAP estimation based on sparse representation ", send out Table is in Signal Processing, the 7th phase in 2014).But, above-mentioned image denoising based on rarefaction representation Method is all to process entire image, does not takes into account the difference of image border and non-edge, and this leads Cause denoising result and often there is the flaw on details and edge, even can lose some texture informations.

As can be seen here, need one badly and can either preferably remove picture noise, and figure can be effectively retained simultaneously As edge and the image de-noising method of detailed information.

Summary of the invention

For the deficiencies in the prior art, it is an object of the invention to provide a kind of figure based on statistics local order feature As denoising method, utilize statistics local order feature as noise constraints condition, it is possible to realize effectively removing image While smooth region noise, it is effectively retained the purpose of image edge details information.Compared with additive method, The method has more preferably denoising effect, it is possible to processing for successive image and analyzing provides the higher denoising of quality Image.

For reaching above-mentioned purpose, the technical solution used in the present invention is as follows:

A kind of image de-noising method based on statistics local order feature, it it is critical only that and follow the steps below:

Step 1: for image I, utilizes local order operator according to formula LRT_k(I)={ LRT_k(I_i)|I_i∈ I} is not Carry out local order conversion under the conditions of same parameter, obtain the positive local order conversion LRT of image_pk(I_i) and negative local order Conversion LRT_nk(I_i), wherein, i is picture numbers, I_iIt is the i-th amplitude and noise acoustic image, k=0, ± 0.01, ± 0.03 ...；

Step 2: described positive local order is converted LRT_pk(I_i) and negative local order conversion LRT_nk(I_i) be added, obtain The statistics local order feature of continuous parameters change

Step 3: based on a kind of image de-noising method, and use described statistics this image that locally order feature draws The constraints of denoising method carries out denoising to image, it is thus achieved that primary picture rich in detail

Step 4: by denoising formula to primary picture rich in detailCarry out second denoising process, and in denoising process The primary picture rich in detail of middle controlWith picture rich in detailBetween statistics local order feature difference, it is achieved noise Remove, it is thus achieved that final picture rich in detail

Further, described image de-noising method is based on crossing complete sparse representation model Denoising Algorithm.

Further, constraints described in step 3 is:

$\underset{{\tilde{I}}_0}{min}\{\underset{k}{\Σ}{E}_k\left(I_i\right)-\underset{k}{\Σ}{E}_k\left({\tilde{I}}_{0i}\right)\},$

Wherein,For noise image I_iStatistics local order feature,For primary clear after filtering Clear imageStatistics local order feature, i is picture numbers, k=0, ± 0.01, ± 0.03 ....

Further, the denoising formula described in step 4 is:

$\underset{\tilde{I}}{min}\{\left(\underset{k}{\Σ}{\mathrm{LRT}}_{pk}\right({\tilde{I}}_i)-\underset{k}{\Σ}{\mathrm{LRT}}_{pk}({\tilde{I}}_{0i}\left)\right)+\left(\underset{k}{\Σ}{\mathrm{LRT}}_{nk}\right({\tilde{I}}_i)-\underset{k}{\Σ}{\mathrm{LRT}}_{nk}({\tilde{I}}_{0i}\left)\right)\}$

$\begin{array}{cc}s.t.& I_N^0=D\mathrm{\α}\end{array}$

Wherein,For picture rich in detailJust statistics local order feature,For clear figure PictureNegative statistics local order feature,For primary picture rich in detailJust statistics local order feature,For primary picture rich in detailNegative statistics local order feature,For primary condition, D was complete Standby dictionary, α is sparse coefficient.

The present invention utilizes local order operator, image carries out under the conditions of different parameters local order conversion, obtains The positive local order conversion of image and the order conversion of negative local；By the conversion of described positive local order and negative local order conversion phase Add the statistics local order feature obtaining continuous parameters change；In traditional image denoising side based on rarefaction representation On the basis of method, using described statistics local order feature as constraints denoising first to image, and remain with figure The edge detail information of picture；By controlling before denoising and described statistics local order feature between image after denoising Difference, carries out second denoising to the image remaining with edge detail information, it is achieved the removal of noise obtains required Picture rich in detail.The present invention utilizes statistics local order feature as noise constraints condition, it is possible to realize removing image While smooth region noise, more effectively retain the purpose of image edge details information.

The remarkable result of the present invention is: utilize statistics local order feature as noise constraints condition, it is possible to realize While to the denoising of image smoothing region, more effectively retain the purpose of image edge details information.With biography The denoising method based on rarefaction representation of system is compared, and the inventive method has more preferably denoising effect, it is possible to obtain Obtain quality higher denoising image, thus ensure the validity and reliability that successive image processes and analyzes.

Accompanying drawing explanation

Fig. 1 is the method flow diagram of the present invention.

Detailed description of the invention

Detailed description of the invention and operation principle to the present invention are described in further detail below in conjunction with the accompanying drawings.

As it is shown in figure 1, a kind of image de-noising method based on statistics local order feature, enter according to following steps OK:

Initially enter step 1: for image I, utilize local order operator according to formula LRT_k(I)={ LRT_k(I_i)|I_i∈ I} carries out local order conversion under the conditions of different parameters, obtains the positive local of image Order conversion LRT_pk(I_i) and negative local order conversion LRT_nk(I_i)；

In the specific implementation, it is possible to use following local order transform operator pair image carries out local order and converts:

LRT_δ(I)={ LRT_δ(I_i)|I_i∈ I},

Wherein,

${\mathrm{LRT}}_{\mathrm{\δ}}\left(I_i\right)=\underset{j}{\Σ}{\#}_{\mathrm{\δ}}(I_i-I_j)\}$

${\#}_{\mathrm{\δ}}(\·)=\left\{\begin{array}{cc}1,& I_i-I_j>\mathrm{\δ}\\ 0,& I_i-I_j\≤\mathrm{\δ}\end{array}\right.$

In formula, δ=kI_avg, k is constant constant, I_avgFor the average of pixel, threshold value in each local neighborhood δ can be on the occasion of or negative value, the corresponding different transformation results of different values.Now the local order conversion of image is as follows:

LRT_k(I)={ LRT_k(I_i)|I_i∈ I},

For noise image I_i, its positive local order conversion is as follows:

LRT_pk(I_i), k=0 ,+0.01 ,+0.03 ...；

Its negative local order conversion is as follows:

LRT_nk(I_i), k=0 ,-0.01 ,-0.03 ...

Enter step 2: in order to image border and non-edge are carried out denoising simultaneously, by described positive local order Conversion LRT_pk(I_i) and negative local order conversion LRT_nk(I_i) be added, the statistics local order obtaining continuous parameters change is special Levy

Secondly entrance step 3: image de-noising method based on rarefaction representation, and use described statistics local order special The constraints of this image de-noising method obtained out carries out denoising to image, it is thus achieved that primary picture rich in detailSpecific as follows:

First noise image rarefaction representation coefficient under crossing complete dictionary D is tried to achieve, as follows:

$\mathrm{\α}={\mathrm{argminI}}_N-{\mathrm{D\α}}_2^2+{\mathrm{\λ\α}}_1;$

Therefore, picture rich in detailAcquisition calculated as below,Wherein, α is rarefaction representation coefficient, and λ uses Control degree of rarefication.Cross complete dictionary D and can use fixing dictionary, it would however also be possible to employ learning method obtains.

Owing to this traditional denoising model based on rarefaction representation is not by edge and the non-edge of image Separately discuss, it is possible that the phenomenon of loss in detail while removing noise, in order to avoid this phenomenon Generation, present invention introduces statistics local order feature constraint item, it is achieved denoising.Described statistics local order feature Bound term is expressed as follows:

$\underset{{\tilde{I}}_0}{min}\{\underset{k}{\Σ}{E}_k\left(I_i\right)-\underset{k}{\Σ}{E}_k\left({\tilde{I}}_{0i}\right)\}$

Wherein,For noise image I_iStatistics local order feature,For primary clear after filtering Clear imageStatistics local order feature, i is picture numbers, k=0, ± 0.01, ± 0.03 ....

Described statistics local order feature constraint item is intended to the statistics of the positive and negative local sum of ranks value so that filtered image Feature is infinitely close to the statistical nature of original image positive and negative local sum of ranks value.

Finally enter step 4: by denoising formula to primary picture rich in detailCarry out second denoising process, and Primary picture rich in detail is controlled during denoisingWith picture rich in detailBetween the difference of statistics local order feature, real The removal of existing noise, it is thus achieved that final picture rich in detail

Described denoising formula is:

$\underset{\tilde{I}}{min}\{\left(\underset{k}{\Σ}{\mathrm{LRT}}_{pk}\right({\tilde{I}}_i)-\underset{k}{\Σ}{\mathrm{LRT}}_{pk}({\tilde{I}}_{0i}\left)\right)+\left(\underset{k}{\Σ}{\mathrm{LRT}}_{nk}\right({\tilde{I}}_i)-\underset{k}{\Σ}{\mathrm{LRT}}_{nk}({\tilde{I}}_{0i}\left)\right)\}$

$\begin{array}{cc}s.t.& I_N^0=D\mathrm{\α}\end{array}$

Wherein,For picture rich in detailJust statistics local order feature,For clear figure PictureNegative statistics local order feature,For primary picture rich in detailJust statistics local order feature,For primary picture rich in detailNegative statistics local order feature,For primary condition, D was complete Standby dictionary, α is sparse coefficient.

For described image de-noising method based on rarefaction representation, use based on crossing complete sparse herein Represent model denoising method, but be not limited to this kind of method.

This programme, will statistics local order feature on the basis of traditional image de-noising method based on rarefaction representation As constraints, image is carried out first denoising, reach to retain the purpose of image edge details information；Subsequently, By controlling before denoising and the difference of described statistics local order feature between image after denoising, to remaining with edge The image of detailed information carries out second denoising, it is achieved the removal of noise.

Therefore, the present invention utilizes statistics local order feature as the constraints of noise, it is achieved that put down image Slide the removal of noise region and to image border and the reservation of detailed information.With traditional based on rarefaction representation Denoising method compare, the inventive method has more preferably denoising effect, it is possible to obtain the higher denoising of quality Image, thus the reliability that effective guarantee successive image processes and analyzes.

Claims (4)

1. an image de-noising method based on statistics local order feature, it is characterised in that enter according to following steps OK:

Step 1: for image I, utilizes local order operator according to formula LRT_k(I)={ LRT_k(I_i)|I_i∈ I} is not Carry out local order conversion under the conditions of same parameter, obtain the positive local order conversion LRT of image_pk(I_i) and negative local order Conversion LRT_nk(I_i), wherein, i is picture numbers, I_iIt is the i-th amplitude and noise acoustic image, k=0, ± 0.01, ± 0.03 ...；

Step 2: described positive local order is converted LRT_pk(I_i) and negative local order conversion LRT_nk(I_i) be added, obtain The statistics local order feature of continuous parameters change

Step 3: based on a kind of image de-noising method, and use described statistics this image that locally order feature draws The constraints of denoising method carries out denoising to image, it is thus achieved that primary picture rich in detail

Step 4: by denoising formula to primary picture rich in detailCarry out second denoising process, and in denoising process The primary picture rich in detail of middle controlWith picture rich in detailBetween statistics local order feature difference, it is achieved noise Remove, it is thus achieved that final picture rich in detail

Image de-noising method based on statistics local order feature the most according to claim 1, its feature exists In: described image de-noising method is based on crossing complete sparse representation model Denoising Algorithm.

Image de-noising method based on statistics local order feature the most according to claim 2, its feature exists In: constraints described in step 3 is:

$\underset{{\tilde{I}}_0}{min}\{\underset{k}{\Σ}{E}_k\left(I_i\right)-\underset{k}{\Σ}{E}_k\left({\tilde{I}}_{0i}\right)\},$

Wherein,For noise image I_iStatistics local order feature,For primary clear after filtering Clear imageStatistics local order feature, i is picture numbers, k=0, ± 0.01, ± 0.03 ....

Image de-noising method based on statistics local order feature the most according to claim 3, its feature exists In: the denoising formula described in step 4 is:

$\underset{\tilde{I}}{\min }\left\{\left(\underset{k}{\mathrm{\Σ}}{\mathrm{LRT}}_{pk}\left({\tilde{I}}_i\right)-\underset{k}{\mathrm{\Σ}}{\mathrm{LRT}}_{pk}\left({\tilde{I}}_{0i}\right)\right)+\left(\underset{k}{\mathrm{\Σ}}{\mathrm{LRT}}_{nk}\left({\tilde{I}}_i\right)-\underset{k}{\mathrm{\Σ}}{\mathrm{LRT}}_{nk}\left({\tilde{I}}_{0i}\right)\right)\right\}$

$\begin{array}{cc}s.t.& I_N^0=D\mathrm{\α}\end{array}$

Wherein,For picture rich in detailJust statistics local order feature,For clear figure PictureNegative statistics local order feature,For primary picture rich in detailJust statistics local order feature,For primary picture rich in detailNegative statistics local order feature,For primary condition, D was complete Standby dictionary, α is sparse coefficient.