CN102045514B - Image noise filtering method - Google Patents

Abstract

The invention discloses an image noise filtering method, which comprises the following steps: sequentially selecting a pixel in the image as a current pixel; dynamically determining a current search frame and an intensity parameter; converting the comparison frame of each pixel in the current search frame to a frequency domain; determining a current frequency reference; calculating the similarity between each neighborhood pixel and the current pixel in the current search frame according to the current frequency reference; determining the weight of each neighborhood pixel relative to the current pixel according to the distance, the similarity and the intensity parameter of each neighborhood pixel and the current pixel in the current search frame; and carrying out weighted average on the gray-scale values of each neighborhood pixel and the current pixel in the current search frame according to the weight so as to obtain the reconstruction value of the current pixel. The method of the invention can determine the noise filtering intensity parameter, the size of the current searching frame and the size of the comparison frame according to the frequency parameter concentration in the frequency domain comparison frame, and can keep more image details.

Description

The picture noise filtering method

Technical field

The present invention relates to a kind of image processing method, and relate in particular to a kind of picture noise filtering method.

Background technology

Picture noise is one of key factor that influences image quality.Yet when the number of pixels of image sensor increased gradually, pixel size was but constantly dwindled because of cost consideration, and the noise that causes being comprised in the image of image sensor acquisition is exaggerated unavoidablely.Therefore, the effect of noise filtering (denoising) becomes a key factor of decision image quality gradually.

Utilize filter (filter) image before the denoising (noisy image) to be redeveloped into the process of image (denoised image) is called image reconstruction (image reconstruction) behind the denoising, as shown in Figure 1.It is understandable that image reconstruction is to carry out through processing unit, and this processing unit generally couples storage element, it is used for the various information that temporary image reconstruction process is produced.

Utilizing neighborhood filter (neighborhood filter) to carry out image reconstruction is a standard techniques.The neighborhood filter is according to the decision of the similarity between current pixel (current pixel) and the neighborhood territory pixel (neighborhood pixel) thereof weight (weighting); And current pixel and its neighborhood territory pixel are carried out weighted average according to this weight, to obtain the reconstructed value of current pixel.After all pixels all execute the step of above-mentioned image reconstruction in the image before the denoising, then can obtain image behind the denoising.Neighborhood filtering generally can be represented with formula (1):

$\hat{U}\left(x\right)=\frac{1}{N_h\left(x\right)}\underset{R_x}{\∫}\∫h(x,y)U\left(y\right)\mathrm{dy}---\left(1\right)$

Wherein, U representes the preceding image of denoising; N _h(x) expression normaliztion constant (normalizationconstant);

Image behind the expression denoising; R _xThe neighborhood of expression current pixel x; H representes filter constant, and it is decided by current pixel x and the distance of its neighborhood territory pixel y in image, the for example position of neighborhood territory pixel y and current pixel x distance (distance), and luminance difference (intensity difference).Please with reference to shown in Figure 2, it shows the sketch map of 7 * 7 neighborhood filtering.Image sensor acquisition image I, it is an image before the denoising.The neighborhood filter is then according to the search frame R around current pixel x and this current pixel x _xSimilarity between the interior neighborhood territory pixel y is obtained 48 weights respectively, and according to these weights the GTG value (gray level) of current pixel x is carried out weighted average with the GTG value of its 48 neighborhood territory pixel y, to obtain the reconstructed value of current pixel x.Yet therefore the neighborhood filter often can't reach gratifying reconstruction effect owing to only merely carry out weighted average according to the similarity between two pixels.

Therefore, other proposes a kind of image reconstruction method, is referred to as non-regional algorithm (non-localalgorithm), is used to improve above-mentioned image reconstruction method based on neighborhood filtering.Non-regional algorithm is main be according to the neighborhood territory pixel of the current pixel comparison block (comparison block) of pre-set dimension around the current pixel and current pixel on every side the similarity between the neighborhood territory pixel comparison block of pre-set dimension decide weight; And the GTG value of current pixel and the GTG value of its neighborhood territory pixel are carried out weighted average according to this weight, to obtain the reconstructed value of current pixel.Non-regional algorithm generally can be represented with formula (2):

$\mathrm{NL}\left[v\right]\left(i\right)=\underset{j\∈\mathrm{Rx}}{\mathrm{\Σ}}\mathrm{\ω}(i,j)v\left(j\right)---\left(2\right)$

Wherein NL [v] (i) representes the reconstructed value of current pixel i; The GTG value of the neighborhood territory pixel j of current pixel i before v (j) the expression denoising; (its current pixel comparison block and neighborhood territory pixel j that is decided by pre-set dimension around the current pixel i is the similarity between the neighborhood territory pixel comparison block of pre-set dimension on every side for i, the j) weight between expression current pixel i and its neighborhood territory pixel j for ω.This weight can be represented with formula (3):

$\mathrm{\ω}(i,j)=\frac{1}{Z\left(i\right)}{e}^{-\frac{{\left|\right|v\left(N_i\right)-v\left(N_j\right)\left|\right|}_{2,a}^2}{h^2}}---\left(3\right)$

Its Chinese style (3) mainly be relative position pixel gray level value between the neighborhood territory pixel comparison block of current pixel comparison block and preset size around the neighborhood territory pixel j of preset size around the expression current pixel i squared difference and; Z (i) then is a normaliztion constant.

For example please with reference to shown in Figure 3, it has shown the sketch map of the non-regional algorithm of utilizing 7 * 7 search frame Rx and 5 * 5 comparison block (Ni, Nj); Wherein, image I is an image before the denoising that image sensor captured; I is a current pixel; Ni is the current pixel comparison block of pre-set dimension (5 * 5) around this current pixel; J is the neighborhood territory pixel of current pixel i; Nj is the neighborhood territory pixel comparison block of pre-set dimension around this neighborhood territory pixel j; Rx is for searching frame.According to shown in Figure 3, the quadratic sum that the weight between current pixel i and the neighborhood territory pixel j is subtracted each other 25 differences of back gained by relative position pixel among current pixel comparison block Ni and the neighborhood territory pixel comparison block Nj determines.Therefore, in searching frame Rx, can try to achieve 48 weights altogether.The reconstructed value of current pixel i is then for to carry out the weighted average gained according to these weights with the GTG value of current pixel i and its neighborhood territory pixel j.

Compared to neighborhood filtering, though above-mentioned non-regional algorithm can obtain preferable denoising effect, because pixel can receive noise effect, so the result who directly the GTG value of two comparison block is carried out computing still can't get rid of the image of noise fully.Therefore, this area has proposed a kind of elder generation in addition with after employed comparison block is converted to a frequency domain in the non-regional algorithm, the method that compares again.Usually in frequency domain, belong to the notion of radio-frequency component based on noise, will be converted to earlier and compare again after part radio-frequency component in the comparison block of frequency domain is removed, to improve the effect of noise filtering.Yet said method also can't dynamically carry out parameter adjustment according to the characteristic of each pixel, causes the details preservation effect relatively poor, and occurs ghost (shock effect) and sawtoothization phenomenons such as (staircasting effect) easily.

The detailed content of above-mentioned image reconstruction method can be published in CVPR2005 with reference to people such as Antoni Buades; Title is " A non-local algorithm for image denoising "; And people such as NouraAzzabou is published in ICIP2007; Title is " Image denoising based on adapteddictionary computation ", paper in disclosed content.

Summary of the invention

In view of this; The present invention proposes a kind of picture noise filtering method in addition; Distribution scenario according to the frequency domain comparison block medium frequency coefficient that is converted to frequency domain; Judge near the complexity of the image of current pixel, and dynamically adjust the noise filtering intensity (denoising strength) of searching frame size and current pixel according to this, have existing side effect in the method now with more image details of reservation and elimination.

The objective of the invention is to propose a kind of picture noise filtering method, the size that it can dynamically adjust current search frame and comparison block according near the complexity of the image current pixel can keep more image details.

Another object of the present invention is to propose a kind of picture noise filtering method, it is applicable to various frequency conversion methods.

Another object of the present invention is to propose a kind of picture noise filtering method that can dynamically adjust the noise filtering intensity of each pixel according to image complexity.

The present invention proposes a kind of picture noise filtering method, and this method comprises the following steps: in image, to select in regular turn a pixel as current pixel, and wherein the pixel around this current pixel is the neighborhood territory pixel of this current pixel; Dynamically around current pixel, determine a current search frame and determine an intensity parameters, and in said current search frame, determine a comparison block around each pixel; The comparison block of each pixel in the said current search frame is converted to a frequency domain, to form the frequency domain comparison block; Determine the current frequency reference of this frequency domain comparison block; Obtain the similarity of each neighborhood territory pixel and current pixel in the said current search frame according to this current frequency reference; According to each neighborhood territory pixel and the distance of current pixel, said similarity and said intensity parameters in the said current search frame, determine the weight of each neighborhood territory pixel with respect to current pixel; And the GTG value of each neighborhood territory pixel and current pixel in the said current search frame is carried out weighted average according to this weight, to obtain the reconstructed value of current pixel.

In picture noise filtering method of the present invention; Dynamically determine a current search frame around the said current pixel and determine a kind of embodiment of an intensity parameters to comprise the following steps: around said current pixel, to determine a maximum frame of searching, wherein should maximum search each pixel in the frame around decision one comparison block is arranged; Calculate said maximum frequency parameter concentration degree of searching frame; And determine said current search frame and said intensity parameters according to this frequency parameter concentration degree.

In picture noise filtering method of the present invention, a kind of embodiment of calculated rate parameter set moderate comprises the following steps: said maximum comparison block of searching all pixels in the frame is converted to a frequency domain, to form the frequency domain comparison block; Calculate said maximum all frequency domain comparison block of searching frame same frequency energy with among, energy and the energy of sum and all frequencies and the merchant of sum of energy and maximum preset number frequency; And the merchant of this energy and sum compared with threshold value, to determine said frequency parameter concentration degree.

The present invention proposes a kind of picture noise filtering method in addition, and this method comprises the following steps: in image, to select in regular turn a pixel as current pixel, and wherein the pixel around this current pixel is the neighborhood territory pixel of this current pixel; Determine a maximum frame of searching around the current pixel, and this maximum search each pixel in frame around determine a comparison block; Said maximum comparison block of searching all pixels in the frame is converted to a frequency domain, to form the frequency domain comparison block; Calculate said maximum edge pixel ratio of searching frame; Determine a current search frame and an intensity parameters according to this edge pixel ratio, and determine the current frequency reference in the said current search frame; Obtain the similarity of each neighborhood territory pixel and current pixel in the said current search frame according to this current frequency reference; According to each neighborhood territory pixel and the distance of current pixel, said similarity and said intensity parameters in the said current search frame, determine the weight of each neighborhood territory pixel with respect to current pixel; And the GTG value of each neighborhood territory pixel and current pixel in the said current search frame is carried out weighted average according to this weight, to obtain the reconstructed value of current pixel.

In picture noise filtering method of the present invention; Calculate said maximum step of searching the edge pixel ratio of frame comprise the following steps: to calculate said maximum all frequency domain comparison block of searching frame same frequency energy with among, energy and the energy of sum and all frequencies and the merchant of sum of energy and maximum preset number frequency; And the merchant of this energy and sum compared with threshold value, to determine said edge pixel ratio

In picture noise filtering method of the present invention; Said similarity is decided by in the frequency domain comparison block of frequency domain comparison block and current pixel of each neighborhood territory pixel in the said current search frame, the energy difference absolute value sum of the energy of each current reference frequency or energy difference quadratic sum.

In picture noise filtering method of the present invention, the energy of the same frequency that said current frequency reference is all frequency domain comparison block with among, energy and maximum preset number frequency.

In the picture noise filtering method of the present invention, for example, but be not limited to, the frequency domain conversion is carried out in discrete cosine transform capable of using, fourier transform, wavelet conversion or principal vector analysis.Method of the present invention can decide the size of noise filtering intensity parameters, current search frame and the size of comparison block according to frequency domain comparison block medium frequency parameter set moderate.

Description of drawings

Fig. 1 has shown the sketch map of image reconstruction;

Fig. 2 has shown the sketch map of existing neighborhood filtering;

Fig. 3 has shown the sketch map of existing non-regional algorithm;

Fig. 4 a has shown the flow chart of the picture noise filtering method of an embodiment of the present invention;

Fig. 4 b has shown the flow chart that determines a current search frame and an intensity parameters among Fig. 4 a;

Fig. 4 c has shown the flow chart of Fig. 4 b calculating frequency parameter set moderate;

Fig. 5 has shown the sketch map of picture noise filtering method of the present invention;

Fig. 6 has shown the sketch map of each frequency that picture noise filtering method frequency domain comparison block of the present invention is comprised;

Fig. 7 a has shown the flow chart of the picture noise filtering method of the another kind of embodiment of the present invention; And

Fig. 7 b has shown the flow chart of decision edge pixel ratio among Fig. 7 a.

The main element symbol description

Image P, P before the I denoising ₁₁～P ₇₇Pixel

The current search frame of Pc current pixel Sc

The maximum frame S that searches of Sc_max ₁～S ₉, A ₁～A ₃Step

The frequency domain comparison block of the comparison block Bpc ' current pixel of Bpc current pixel

B _P11～B _P77Comparison block B _P11'～B _P77' frequency domain comparison block

E ₁～E ₂₅The energy of each frequency in the frequency domain comparison block

E ₁ ^Sum～E ₂₅ ^SumIn the frequency domain comparison block of current search frame the energy of each frequency with

Rx searches frame x, y pixel

I, j pixel Ni, Nj comparison block

Embodiment

For let above-mentioned and other purposes of the present invention, feature and advantage can be more obvious, hereinafter will cooperate appended diagram, elaborate as follows.In addition, need to prove that in explanation of the present invention, identical member is with identical symbolic representation.

Please with reference to shown in Fig. 4 a; It has shown the picture noise filtering method according to an embodiment of the present invention; This method comprises the following steps: in image, to select in regular turn a pixel as current pixel, and wherein the pixel around this current pixel is neighborhood territory pixel (the step S of this current pixel ₁); Dynamically around current pixel, determine a current search frame and determine an intensity parameters, and in said current search frame, determine a comparison block (step S around each pixel ₂); The comparison block of each pixel in the current search frame is converted to a frequency domain, to form frequency domain comparison block (step S ₃); Determine current frequency reference (the step S of said frequency domain comparison block ₄); Obtain similarity (the step S of each neighborhood territory pixel and current pixel in the current search frame according to this current frequency reference ₅); According to distance, said similarity and the said intensity parameters of each neighborhood territory pixel and current pixel in the current search frame, determine weight (the step S of each neighborhood territory pixel with respect to current pixel ₆); According to this weight the GTG value of each neighborhood territory pixel and current pixel in the current search frame is carried out weighted average, to obtain reconstructed value (the step S of current pixel ₇); And all pixels that judge whether said image have all been tried to achieve reconstructed value (step S ₈); If accomplish reconstruction (the step S of said image ₉); If not, execution in step S again then ₁

Please with reference to shown in Fig. 4 b, it has shown the step S of Fig. 4 a ₂In dynamically determine a kind of embodiment of a current search frame and an intensity parameters, comprise the following steps: around current pixel, to determine a maximum frame of searching, wherein should maximum search each pixel in the frame around decision one comparison block (step S is arranged ₂₁); Calculate maximum frequency parameter concentration degree (the step S that searches frame ₂₂); And determine current search frame and intensity parameters (step S according to this frequency parameter concentration degree ₂₃).

Please with reference to shown in Fig. 4 c, it has shown the step S of Fig. 4 b ₂₂A kind of embodiment of calculating frequency parameter set moderate comprises the following steps: maximum searched that the comparison block of all pixels is converted to a frequency domain in the frame, to form frequency domain comparison block (step S ₂₂₁); Calculate maximum all frequency domain comparison block of searching frame same frequency energy with among, energy and the energy of sum and all frequencies and merchant (the step S of sum of energy and maximum preset number frequency ₂₂₂); And the merchant of this energy and sum compared with threshold value, to determine said frequency parameter concentration degree (step S ₂₂₃).Wherein, said threshold value is a numerical value, its big I according to the noise filtering effect desiring to reach decide.

The detailed execution mode of the picture noise filtering method of the embodiment of the invention then is described; Picture noise filtering method of the present invention carries out image reconstruction to form image (denoised image) behind the denoising to image (noisy image) before the denoising, and is as shown in Figure 1.It is understandable that above-mentioned each step is carried out through processing unit, and this processing unit couples memory cell, with the various information that produced in the access image reconstruction process.

Please with reference to shown in Figure 5, image I comprises the pixel P that a plurality of matrix forms are arranged before the denoising, and each pixel P has a GTG value (gray level), and wherein the size of image I can determine according to practical application.Picture noise filtering method of the present invention is obtained before the denoising reconstruction GTG value (reconstructed gray level) of all pixel P in the image I, and according to the reconstruction GTG value of being obtained to form image behind the denoising.

Please be simultaneously with reference to Fig. 4 a and shown in Figure 5; Picture noise filtering method of the present invention for example; But be not limited to, select to begin to calculate the reconstruction GTG value of this pixel, obtain before the denoising reconstruction GTG value of other pixels in the image I then in regular turn from first pixel in a corner of image I.In explanation of the present invention, the current pixel of handling is called current pixel (current pixel) Pc, and the pixel around this current pixel Pc is the neighborhood territory pixel of this current pixel Pc, for example P ₁₁, P ₁₂..., P ₇₇(step S ₁).

Dynamically around current pixel Pc, determine a current search frame Sc and determine an intensity parameters (step S ₂).In the present invention, the size of current search frame Sc is that the complexity according near the image current pixel Pc is dynamically determined, complexity is higher, and the size of current search frame Sc is littler; Complexity is lower, and the size of current search frame Sc is bigger.The present invention is through selecting different current search frame Sc sizes to improve the effect of image denoising sound.Said intensity parameters then is used for subsequent step, is used to determine image denoising sound intensity (being specified in the back).At step S ₂In, determine the comparison block B around all pixel P, for example pixel P in the current search frame Sc simultaneously ₁₁Comparison block B _P11..., pixel P ₇₇Comparison block B _P77And comparison block B _P11～B _P77Size for example be 5 * 5.Therefore, if current search frame Sc is of a size of 7 * 7, then it comprises 49 pixel P ₁₁～P ₇₇And Pc, determine to have corresponding comparison block B around these pixels respectively _P11～B _P77And Bpc.It is understandable that the size of above-mentioned current search frame Sc and comparison block is not limited to content disclosed herein.

Please be simultaneously with reference to Fig. 4 b and Fig. 5, then explanation determines current search frame Sc size and the strong and weak a kind of embodiment of decision intensity parameters.(step S after current pixel Pc is determined ₁), at first around current pixel Pc, determining a maximum frame Sc_max that searches, its size for example is 7 * 7 (step S ₂₁), and should maximum search among the frame Sc_max and determined that all a comparison block, its size for example are 5 * 5 around each pixel.For example, pixel P ₁₁Decision on every side has a comparison block B _P11, decision has a comparison block Bpc around the current pixel Pc.Then, the comparison block of maximum being searched all pixels among the frame Sc_max is converted to a frequency domain, to form the frequency domain comparison block.For example, comparison block B _P11The conversion back forms B _P11', comparison block Bpc conversion back forms Bpc '; Other 47 comparison block also all are converted into the frequency domain comparison block among the in like manner maximum search frame Sc_max.The mode of frequency inverted for example but is not limited to, and can use discrete cosine transform, fourier transform, wavelet conversion or principal vector analysis.

Then, calculate maximum frequency parameter concentration degree (the step S that searches frame ₂₂), after its execution mode will be specified in.In the present invention, maximum meaning of searching the frequency parameter concentration degree of frame is equivalent to maximum edge pixel ratio of searching among the frame Sc_max.Judge that maximum whether search a pixel in the frame be that the mode of edge pixel for example does, after the comparison block of this pixel converts the frequency domain comparison block into, if the concentration of energy of this frequency domain comparison block representes then that in some CF this pixel belongs to edge pixel; Otherwise,, represent that then this pixel does not belong to edge pixel if the energy even of this frequency domain comparison block is distributed in all frequencies.Therefore, the frequency parameter concentration degree of searching frame when maximum is higher, representes that then this maximum search frame comprises more edge pixel.Then, decide power (the step S of size and the intensity parameters of current search frame Sc according to the frequency parameter concentration degree ₂₃).For example, when frequency parameter concentration degree (or edge pixel ratio) was higher, the image in the image range of the maximum search of expression frame Sc_max was complicated, therefore selects less current search frame Sc and more weak intensity parameters; Otherwise when frequency parameter concentration degree (or edge pixel ratio) was low, the image in the image range of the maximum search of expression frame Sc_max was milder, therefore selects bigger current search frame Sc and stronger intensity parameters.In the present invention, according to the frequency parameter concentration degree, the size of current search frame Sc for example can be 7 * 7,5 * 5 or 3 * 3.Yet the present invention is not limited to above-mentioned disclosed content, and spirit of the present invention is and can decides the size of current search frame Sc and the power of intensity parameters according to the complexity of image around the current pixel Pc.

Please be simultaneously with reference to Fig. 4 c to Fig. 6, then explanation determines a kind of embodiment of said frequency parameter concentration degree.In this explanation, suppose that the maximum frame Sc_max that searches is of a size of 7 * 7, the comparison block of each pixel is of a size of 5 * 5., a comparison block comprises 25 frequencies, for example pixel P after being converted into a frequency domain comparison block ₁₁Comparison block B _P11Form B after being converted to frequency domain _P11', this frequency domain comparison block B _P11' comprise 25 frequencies (for example frequency 1～25, wherein 1～25 is merely numbering and is not to represent actual frequency), and each frequency has an energy E respectively ₁ ¹¹～E ₂₅ ¹¹The frequency domain comparison block Bpc ' of current pixel Pc comprises the energy E of 25 frequencies ₁ ^Pc～E ₂₅ ^PcFrequency domain comparison block B _P77' comprise the energy E of 25 frequencies ₁ ⁷⁷～E ₂₅ ⁷⁷Be convenient explanation, among Fig. 6 each frequency domain comparison block be shown as one 25 array of tieing up (array), wherein the subscript of energy E is represented the frequency numbering, and subscript is remarked pixel (step S then ₂₂₁).

With said frequency domain comparison block (B _P11'～B _P77') in the energy addition of same frequency, in the hope of the energy and the E of each frequency ₁ ^Sum～E ₂₅ ^Sum, the energy of frequency 1 and do for example $E_{}^{}$ ${}_1^{\mathrm{Sum}}={\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="H2009101804038E00021.png,H2009101804038E00022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1^{11}+...+E_1^{77},...,$ The energy of frequency 25 and do ${\mathrm{<figure-callout\; id="25"\; label="E"\; filenames="H2009101804038E00061.png"\; state="\{\{state\}\}">E</figure-callout>}}_{25}^{\mathrm{Sum}}={\mathrm{<figure-callout\; id="25"\; label="E"\; filenames="H2009101804038E00061.png"\; state="\{\{state\}\}">E</figure-callout>}}_{25}^{11}+...+{\mathrm{<figure-callout\; id="25"\; label="E"\; filenames="H2009101804038E00061.png"\; state="\{\{state\}\}">E</figure-callout>}}_{25}^{77}.$ Then, with the energy and the addition of energy and maximum preset number frequency, and divided by each frequency energy and sum merchant (the step S in the hope of an energy and sum ₂₂₂).For example suppose that energy and 5 maximum frequencies are E ₁ ^Sum, E ₃ ^Sum, E ₅ ^Sum, E ₇ ^SumAnd E ₉ ^Sum, then the merchant of energy and sum can be expressed as (E ₁ ^Sum+ E ₃ ^Sum+ E ₅ ^Sum+ E ₇ ^Sum+ E ₉ ^Sum)/(E ₁ ^Sum+ ... + E ₂₅ ^Sum).

Then, the merchant of energy of being obtained and sum is compared with threshold value, and, determine said frequency parameter concentration degree (step S according to the merchant of this energy and sum and the relation of threshold value ₂₂₃).For example as the merchant of energy and sum during greater than threshold value, the maximum image of searching in the frame Sc_max image range of expression is complicated; Otherwise as the merchant of energy and sum during less than threshold value, the maximum image of searching in the frame Sc_max image range of expression is milder.It is understandable that, judge that the account form of frequency parameter concentration degree is not defined as content disclosed herein, for example also can utilize other statistical methods to calculate the distribution scenario of each frequency energy.It is understandable that in addition the comparison block size of each pixel also can dynamically determine according to frequency parameter concentration degree (edge pixel ratio) among the current search frame Sc.

As step S ₂After the completion, then can determine a current search frame Sc and an intensity parameters according near the complexity of the image current pixel Pc.Then, the comparison block of each pixel among the current search frame Sc is converted to a frequency domain, to form frequency domain comparison block (step S ₃), be converted to the same discrete cosine transform capable of using of mode, fourier transform, wavelet conversion or the principal vector analysis of frequency domain here.It is understandable that, if at step S ₂In maximum has been searched each pixel among the frame Sc_max comparison block convert the frequency domain comparison block into, then can earlier these frequency domain comparison block be stored in storage element.Because current search frame Sc searches frame Sc_max smaller or equal to maximum, therefore at step S ₃Then can be directly read the current search frame Sc with respect to the frequency domain comparison block of each pixel and the frequency domain that need not try again conversion from storage element.

Please refer again to Fig. 4 a, Fig. 5 and shown in Figure 6; As current search frame Sc; Its size for example is 7 * 7; The comparison block of each pixel be converted into the frequency domain comparison block after, its size for example is 5 * 5, then knack search the relevant frequency domain comparison block of frame Sc before settled current frequency reference (frequency basis) for using in the subsequent step.Please refer again to Fig. 6, in the present invention, the current frequency reference decision of current search frame Sc is the energy of each frequency and E ₁ ^Sum～E ₂₅ ^SumAmong, energy and maximum preset number frequency.Because the energy of a frequency is bigger with, represent that then this frequency has comprised most image information among the current search frame Sc.For example, suppose E ₁₀ ^Sum～E ₁₀ ^SumBeing energy and 10 maximum frequencies, is current frequency reference (step S with this 10 frequency (frequency numbering 10～19) then ₄).

Then, obtain similarity (the step S of each neighborhood territory pixel and current pixel in the current search frame according to current frequency reference ₅).Account form is to calculate among the current search frame Sc, belongs to the energy absolute difference sum or the energy difference quadratic sum that belong to each frequency energy of current frequency reference in the frequency domain comparison block of each frequency energy and current pixel Pc of this current frequency reference in the frequency domain comparison block of each neighborhood territory pixel of current pixel Pc.When the value of this energy absolute difference sum or energy difference quadratic sum more hour, then represent the similarity height of a neighborhood territory pixel and current pixel; Otherwise,, represent that then the similarity of a neighborhood territory pixel and current pixel is low when the value of energy absolute difference sum or energy difference quadratic sum heals when big.Pixel P for example ₁₁Can determine according to following formula with the similarity of current pixel Pc: (| E ₁₀ ¹¹-E ₁₀ ^Pc|+| E ₁₁ ¹¹-E ₁₁ ^Pc|+... + | E ₁₉ ¹¹-E ₁₉ ^Pc|) or [(E ₁₀ ¹¹-E ₁₀ ^Pc) ²+ (E ₁₁ ¹¹-E ₁₁ ^Pc) ²+ ... + (E ₁₉ ¹¹-E ₁₉ ^Pc) ²].It is understandable that; One neighborhood territory pixel and current pixel calculation of similarity degree mode are not limited to above-mentioned disclosed content, also can utilize other modes to represent the relation of each frequency energy among the frequency domain comparison block Bpc ' of frequency domain comparison block and current pixel Pc of a neighborhood territory pixel.Can obtain the similarity of 48 neighborhood territory pixels and current pixel Pc in this embodiment, current pixel Pc and itself then have hundred-percent similarity.

At this moment, can be according to each neighborhood territory pixel P among the current search frame Sc ₁₁～P ₇₇With distance, similarity and the intensity parameters of current pixel Pc, decision pixel P ₁₁～P ₇₇Weight (weighting) (step S with respect to current pixel Pc ₆).When a neighborhood territory pixel is far away apart from current pixel Pc, reduce the weight of this neighborhood territory pixel, when a neighborhood territory pixel is nearer apart from current pixel Pc, increase the weight of this neighborhood territory pixel; When the similarity of a neighborhood territory pixel and current pixel Pc is low, reduce the weight of this neighborhood territory pixel, when the similarity of a neighborhood territory pixel and current pixel Pc is high, increase the weight of this neighborhood territory pixel.All neighborhood territory pixel P of Sc in current search frame ₁₁～P ₇₇After the distance and similarity decision weight proportion according to itself and current pixel Pc, cooperate intensity parameters just can determine each neighborhood territory pixel P ₁₁～P ₇₇Weight with respect to current pixel Pc; Wherein, when image in the current search frame Sc scope is complicated, select lower intensity parameters, to reduce denoising intensity; Otherwise, when image in the current search frame Sc scope is mild, select higher intensity parameters, to increase denoising intensity.Said intensity parameters is used for adjusting said weight proportion according to image complexity in the current search frame Sc scope, so intensity parameters for example can be mathematical function (ratio value, power, log function or other mathematical functions) or numerical value, but is not limited to this; For example intensity parameters is used for weight proportion is multiplied by a ratio value, weight proportion is got power, weight proportion is got the log function or added a numerical value etc.

Each neighborhood territory pixel P in trying to achieve current search frame Sc ₁₁～P ₇₇After the weight of current pixel Pc, then according to this weight with current pixel Pc and each neighborhood territory pixel P ₁₁～P ₇₇The GTG value, for example utilize formula (2), carry out weighted average, to obtain reconstructed value (the step S of current pixel ₆).

Then, judge whether that all pixel P have all tried to achieve reconstructed value (step S in the preceding image I of denoising ₇).When all pixel P have all tried to achieve reconstructed value, then accomplish the reconstruction of image I, and image behind the generation denoising.If do not accomplish the reconstruction of all pixel P, then get back to step S ₁Carry out the image reconstruction of next pixel P in the image I.

Please with reference to shown in Fig. 7 a; It has shown according to the present invention the picture noise filtering method of another kind of embodiment; This method comprises the following steps: in image, to select in regular turn a pixel as current pixel, and wherein the pixel around this current pixel is neighborhood territory pixel (the step S of this current pixel ₁); Determine a maximum frame of searching around the current pixel, and in this maximum search each pixel in frame around determine a comparison block (step S ₂₁); Maximum searched the comparison block of all pixels is converted to a frequency domain in the frame, to form frequency domain comparison block (step S ₂₂₁); Calculate maximum edge pixel ratio (steps A of searching frame ₁); Determine a current search frame and an intensity parameters according to this edge pixel ratio, and determine the current frequency reference (steps A of current search frame ₂); Obtain similarity (the step S of each neighborhood territory pixel and current pixel in the current search frame according to this current frequency reference ₅); According to distance, said similarity and the said intensity parameters of each neighborhood territory pixel and current pixel in the current search frame, determine weight (the step S of each neighborhood territory pixel with respect to current pixel ₆); According to this weight the GTG value of each neighborhood territory pixel and current pixel in the current search frame is carried out weighted average, to obtain reconstructed value (the step S of current pixel ₇); And judge whether that all pixels have all tried to achieve reconstructed value (step S ₈); If, reconstruction (the step S of completion image ₉); If not, execution in step S again then ₁Among this embodiment, the step identical with Fig. 4 a-4c represented with same numeral.In addition, as previously mentioned, maximum edge pixel ratio of searching frame is same as the frequency parameter concentration degree on meaning, so the steps A of Fig. 7 a ₁Be similar to the step S of Fig. 4 b ₂₂Steps A ₂Be similar to step S ₂₃And S ₄The difference of the embodiment of this embodiment and Fig. 4 a-4c only is the enforcement order of step, and execution mode is then similar, because its detailed execution mode can be with reference to aforementioned, so repeat no more in this.

In addition, please with reference to shown in Fig. 7 a and Fig. 7 b, steps A ₁In the maximum mode of searching the edge pixel ratio of frame of decision also comprise the following steps: to calculate maximum all frequency domain comparison block of searching frame same frequency energy with among, energy and the energy of sum and all frequencies and merchant (the step S of sum of energy and maximum preset number frequency ₂₂₂); And the merchant of this energy and sum compared with threshold value, to determine said edge pixel ratio (steps A ₃).Because maximum edge pixel ratio of searching frame is same as the frequency parameter concentration degree, the steps A of Fig. 7 b on meaning ₃Be similar to the step S of Fig. 4 c ₂₂₃

Mandatory declaration be, above-mentioned each form (frame) comprises current search frame, maximum searches frame, frequency domain comparison block, searches frame and comparison block etc., though explain with square, it is not to be used to limit the present invention.Said form (frame) can be arbitrary shape, for example rectangle, rhombus, circle or ellipse etc. according to the actual requirements.

As previously mentioned, because existing image reconstruction method can't dynamically be carried out parameter adjustment according to the characteristic of each pixel, cause the details preservation effect relatively poor.The present invention proposes a kind of picture noise filtering method (Fig. 4 a-4c and Fig. 7 a-7b) that can dynamically adjust noise filtering intensity, search frame size and the comparison block size of each pixel according to image complexity in addition, can keep more image details and can eliminate existing side effect in the existing method.

Though the present invention is disclosed by the foregoing description, yet the foregoing description is not to be used to limit the present invention, and those skilled in the art under any the present invention are not breaking away from the spirit and scope of the present invention, should do various variations and modification.Therefore protection scope of the present invention should be as the criterion with the scope that appended claims was defined.

Claims (12)

1. picture noise filtering method, this method comprises the following steps:

In image, select a pixel as current pixel in regular turn, wherein the pixel around this current pixel is the neighborhood territory pixel of this current pixel;

Complexity according to image around the said current pixel dynamically determines a current search frame and determines an intensity parameters around said current pixel, and in said current search frame, determines a comparison block around each pixel;

The comparison block of each pixel in the said current search frame is converted to a frequency domain, to form the frequency domain comparison block;

Determine the current frequency reference of said frequency domain comparison block, the energy of the same frequency that wherein said current frequency reference is all frequency domain comparison block with among, energy and maximum preset number frequency;

Obtain the similarity of each neighborhood territory pixel and current pixel in the said current search frame according to this current frequency reference;

According to each neighborhood territory pixel and the distance of current pixel, said similarity and said intensity parameters in the said current search frame, determine the weight of each neighborhood territory pixel with respect to current pixel; And

According to this weight the GTG value of each neighborhood territory pixel and current pixel in the said current search frame is carried out weighted average, to obtain the reconstructed value of current pixel.

2. picture noise filtering method according to claim 1, this method also comprises the following steps:

All pixels that judge whether said image have all been tried to achieve reconstructed value;

Wherein, the step that the comparison block of each pixel in the said current search frame is converted to a frequency domain utilizes discrete cosine transform, fourier transform, wavelet conversion or principal vector analysis to realize;

Wherein, the energy of the same frequency that said current frequency reference is all frequency domain comparison block with among, energy and maximum preset number frequency.

3. picture noise filtering method according to claim 1; Wherein, said similarity is the energy difference absolute value sum or the energy difference quadratic sum of the energy of each current reference frequency in the frequency domain comparison block of frequency domain comparison block and current pixel of each neighborhood territory pixel in the said current search frame; That said current search frame, comparison block and frequency domain comparison block are respectively done for oneself is square, rectangle, rhombus, circle or oval.

4. picture noise filtering method according to claim 1 wherein, dynamically determines a current search frame and determines the step of an intensity parameters to comprise the following steps: around said current pixel

Determine a maximum frame of searching around the said current pixel, wherein should maximum search each pixel in the frame around decision one comparison block is arranged;

Said maximum comparison block of searching all pixels in the frame is converted to a frequency domain, to form the frequency domain comparison block;

According to the distribution scenario of each frequency energy in the said frequency domain comparison block of said maximum all pixels of searching frame to calculate said maximum frequency parameter concentration degree of searching frame to represent said complexity; And

Determine said current search frame and said intensity parameters according to this frequency parameter concentration degree.

5. picture noise filtering method according to claim 4; Wherein, the distribution scenario according to each frequency energy in the said frequency domain comparison block of said maximum all pixels of searching frame comprises the following steps: to calculate said maximum step of searching the frequency parameter concentration degree of frame

Calculate said maximum all frequency domain comparison block of searching frame same frequency energy with among, energy and the energy of sum and all frequencies and the merchant of sum of energy and maximum preset number frequency; And

The merchant of this energy and sum is compared with threshold value, to determine said frequency parameter concentration degree;

Wherein, the step that said maximum comparison block of searching all pixels in the frame is converted to a frequency domain utilizes discrete cosine transform, fourier transform, wavelet conversion or principal vector analysis to realize;

Wherein, the said maximum frame of searching is of a size of 7 * 7.

6. picture noise filtering method according to claim 4, wherein, said frequency parameter concentration degree is higher, and said the intensity parameters low and said current search frame of healing is littler; Said frequency parameter concentration degree is lower, and said intensity parameters is higher and said current search frame is bigger; According to said frequency parameter concentration degree, said current search frame is of a size of 7 * 7,5 * 5 or 3 * 3.

7. picture noise filtering method, this method comprises the following steps:

In image, select a pixel as current pixel in regular turn, wherein the pixel around this current pixel is the neighborhood territory pixel of this current pixel;

Determine a maximum frame of searching around the said current pixel, and this maximum search each pixel in frame around determine a comparison block;

Said maximum comparison block of searching all pixels in the frame is converted to a frequency domain, to form the frequency domain comparison block;

Calculate said maximum edge pixel ratio of searching frame; Wherein said edge pixel be the concentration of energy of said frequency domain comparison block in some CF but not be uniformly distributed in the said maximum pixel of searching in the frame of all frequencies, said edge pixel ratio is said maximum ratio of searching edge pixel described in the frame and all pixels;

Determine a current search frame and an intensity parameters according to this edge pixel ratio; And determine the current frequency reference of said current search frame; The energy of the same frequency that wherein said current frequency reference is all frequency domain comparison block with among, energy and maximum preset number frequency;

Obtain the similarity of each neighborhood territory pixel and current pixel in the said current search frame according to this current frequency reference;

According to each neighborhood territory pixel and the distance of current pixel, said similarity and said intensity parameters in the said current search frame, determine the weight of each neighborhood territory pixel with respect to current pixel; And

According to this weight the GTG value of each neighborhood territory pixel and current pixel in the said current search frame is carried out weighted average, to obtain the reconstructed value of current pixel.

8. picture noise filtering method according to claim 7, this method also comprises the following steps:

All pixels that judge whether said image have all been tried to achieve reconstructed value;

Wherein, the step that said maximum comparison block of searching all pixels in the frame is converted to a frequency domain utilizes discrete cosine transform, fourier transform, wavelet conversion or principal vector analysis to realize;

Wherein, the energy of the same frequency that said current frequency reference is all frequency domain comparison block with among, energy and maximum preset number frequency.

9. picture noise filtering method according to claim 7 wherein, calculates said maximum step of searching the edge pixel ratio of frame and comprises the following steps:

Calculate said maximum all frequency domain comparison block of searching frame same frequency energy with among, energy and the energy of sum and all frequencies and the merchant of sum of energy and maximum preset number frequency; And

The merchant of this energy and sum is compared with threshold value, to determine said edge pixel ratio.

10. picture noise filtering method according to claim 7, wherein, said edge pixel ratio is higher, and said the intensity parameters low and said current search frame of healing is littler; Said edge pixel ratio is lower, and said intensity parameters is higher and said current search frame is bigger; According to said edge pixel ratio, said current search frame is of a size of 7 * 7,5 * 5 or 3 * 3.

11. picture noise filtering method according to claim 7; Wherein, said similarity is the energy difference absolute value sum or the energy difference quadratic sum of the energy of each current reference frequency in the frequency domain comparison block of frequency domain comparison block and current pixel of each neighborhood territory pixel in the said current search frame; Said maximum frame, comparison block, frequency domain comparison block and current search frame respectively do for oneself square, rectangle, rhombus, the circle or oval of searching.

12. picture noise filtering method according to claim 7, wherein, said intensity parameters is mathematical function or numerical value.

Images