CN100556075C - Remove the method for picture noise - Google Patents

Abstract

The invention discloses a kind of method of removing picture noise, in the process of using bidirectional filter (Bilateralfilter) removal picture noise, the filtering image noise intensity apart from filter in the bidirectional filter is adjusted on the border whether increase has an image according to regional area in the image, and increase the intensity of adjusting the similar filter filtering image noise in the bidirectional filter with the lightness of regional area, make bidirectional filter remove the better effects if of picture noise.

Description

Remove the method for picture noise

Technical field

The present invention relates to a kind of method of removing picture noise, particularly relate to and a kind ofly can adjust the similar weighted value of bidirectional filter and distance weighting value to remove the method for picture noise according to image boundary and lightness.

Background technology

Remove in the method for picture noise at present, the practice of some is to use median filter (MedianFilter), average filter (Mean filter) or low pass filter (Low Pass Filter; LPF) etc. method is removed picture noise, but above-mentioned several method is to use the mode that the pixel in the entire image is averaged to remove picture noise, and not with smooth region in the image (zone that most pixel is similar) and details area (zone that the border is arranged) separate processes, when image has smooth region and details area simultaneously, pixel in the details area can be average because of the pixel in smooth region, and the more and more fuzzyyer phenomenon of details area takes place.Simultaneously, using average mode to remove picture noise often need remove picture noise and keep making choice, the problem that this will cause picture quality to improve between the details area.Therefore the method that removes picture noise with bidirectional filter was suggested 1998 Christian eras.

Bidirectional filter use two filters relevant with distance and similarity come each pixel in the reconstructed image.Be called apart from filter (domain filter) with the filter of distance dependent, the reference value that refers to the near more reference pixel of distance objective pixel is high more, make the time, the closer to the distance weighting value (W of the reference pixel of object pixel according to each reference pixel reconstructed object pixel _Domain) high more; The filter relevant with similarity is similar filter (range filter), refer in each reference pixel around the object pixel, the reference pixel similar more to object pixel has high more reference value, makes according to each reference pixel reconstructed object pixel the time similar weighted value (W to the similar more reference pixel of object pixel _Range) high more, wherein, all be translation invariant Gaussian filter (shift-invariant Gaussian filter) apart from filter and similar filter.As shown in Figure 1, if bidirectional filter uses with the shielding (Mask) as term of reference of 3 * 3 form 100, make that the object pixel 101 of form 100 central authorities will be according to first reference pixel, 102 to the 8th reference pixels 108 on every side and rebuilt, the formula of reconstruction is:

$P_{\mathrm{result}}=\frac{\underset{P_i\∈\mathrm{Mask}}{\mathrm{\Σ}}{P}_i\×W_{\mathrm{range},i}\×W_{\mathrm{domain},i}}{\mathrm{\Σ}{W}_{\mathrm{range},i}\×W_{\mathrm{domain},i}}$

Wherein, i equals 1 to 8, and corresponding first reference pixel, 102 to the 8th reference pixels 108 that is to say P respectively _ResultBe the object pixel 101 after rebuilding; W _{Range, 1}, be the similar weighted value of first reference pixel, 102 target 101; W _{Domain, 1}, be the distance weighting value of first reference pixel, 102 target 101, W _{Range, 2}, W _{Domain, 2}, etc. the rest may be inferred.

Under the increasing situation of the image of present high ISO number, because the image of high ISO number must be accompanied by the hi-vision generating noise, therefore the image of high ISO number often need be removed picture noise.Though use the method for bidirectional filter.Can be when removing noise separate processes smooth region and details area, make picture quality improve, but bidirectional filter promotes the effect of picture quality when removing the hi-vision noise limited.

Summary of the invention

The object of the present invention is to provide a kind of method of removing picture noise, can improve and remove the function that high ISO counts the picture noise of image, by when using bidirectional filter to remove picture noise, adjust the distance weighting value and the similar weighted value of bidirectional filter reconstructed object pixel, make adjusted bidirectional filter remove anti noise and improve, so as to solving the problem that known technology is mentioned.

To achieve these goals, the invention provides a kind of method of removing picture noise, include the following step: one of them is object pixel by selected each not rebuilt pixel in the image that comprises several pixels, and several reference pixels of target setting pixel correspondence; Calculate the similarity value of each reference pixel target, the lightness conversion according to image simultaneously obtains bright deviant, and calculates the similar weighted value of each reference pixel target with bright deviant and similarity value; Calculate the distance weighting value of each reference pixel target according to the distance value of each reference pixel target; Calculate the smoothness value of object pixel, converting according to the smoothness value obtains level and smooth weighted value, and adjusts the distance weighting value with level and smooth weighted value; Similar weighted value and adjusted distance weighting value reconstructed object pixel with each reference pixel target.

Describe the present invention below in conjunction with the drawings and specific embodiments, but not as a limitation of the invention.

Description of drawings

Fig. 1 is the position view of object pixel and reference pixel;

Fig. 2 is the method flow diagram of the removal picture noise that the present invention carried;

Fig. 3 A for brightness that the present invention carried to bright deflection graph;

Fig. 3 B is the adjusted similarity weight multigraph that the present invention carried;

Fig. 3 C is the adjusted similarity weight multigraph that the present invention carried;

Fig. 4 is for the present invention carried apart from weights figure.

Wherein, Reference numeral:

100 forms

101 object pixels

102 to 109 reference pixels

The 310a first similar weighting curve

The 310b second similar weighting curve

The 310c third phase is like weighting curve

311 first bright deviants

312 second bright deviants

The 410a first distance weighting curve

410b second distance weighting curve

Step 210 selected target pixel is also set reference pixel

Step 220 is calculated similar weighted value and is adjusted similar weighted value with lightness

Step 230 is calculated the distance weighting value and is adjusted the distance weighting value with the smoothness value

Step 240 reconstructed object pixel

Embodiment

All is to calculate with translation invariant Gaussian filter (Gauss's smooth function) to produce at bidirectional filter with removing the similar weighted value of noise and distance weighting value, owing to be used for calculating similar weighted value and distance weighting value Gauss smooth function is respectively form

$S(\mathrm{\ξ},x)=e^{-\frac{1}{2}\left(\frac{\mathrm{\δ}(f\left(\mathrm{\ξ}\right),f\left(x\right))}{{\mathrm{\σ}}_r}\right)}C(\mathrm{\ξ},x)=e^{-\frac{1}{2}\left(\frac{d(\mathrm{\ξ},x)}{{\mathrm{\σ}}_d}\right)},$

Wherein ξ is a reference pixel, x is an object pixel, f (x1) is the intensity of pixel x1, δ (f (x1), f (x2)) is the strength difference value of pixel x1 and pixel x2, d (x1, x2) be distance between pixel x1 and pixel x2, σ d is for adjusting the parameter apart from Gaussian function intensity, σ r is for adjusting the parameter of similar Gaussian function intensity, so after in fact can having calculated similar weighted value and distance weighting value earlier, use discrete weighted value table to replace the calculating of Gaussian filter, the mode that present embodiment is also tabled look-up use substitutes the computational process of Gaussian filter.

The present invention be mainly the similar filter adjusted in the bidirectional filter with apart from filter, just adjust the similar weighted value and the distance weighting value of each reference pixel of reconstructed object pixel, making known bidirectional filter remove anti noise (especially removing the effect of strong noise) is obtained to promote, when therefore using the present invention that an image is removed noise, the method and the bidirectional filter of work are similar, all pixels in the same meeting reconstructed image, difference is than the known step of adjusting similar weighted value and distance weighting value that increased.

Below will with an embodiment with explain orally operational system of the present invention and method, and please refer to the method flow diagram of the removal picture noise that Fig. 2 the present invention carried.Select a pixel as object pixel (step 210) in meeting at first of the present invention each pixel by image, after object pixel 101 is chosen, the present invention sets term of reference and is as shown in Figure 13 * 3 nine palace lattice 100, wherein object pixel 101 is in the central authorities of nine palace lattice 101, remaining is the reference pixel of object pixel 101, meeting then of the present invention is calculated the similar weighted value and the distance weighting value of first to the 8th reference pixel, 102 to 109 target 101 according to the method for known bidirectional filter, and adjusts similar weighted value and distance weighting value after calculating.The step of wherein adjusting similar weighted value and distance weighting value does not have the relation of precedence, that is to say, can adjust similar weighted value earlier, can adjust the distance weighting value earlier yet.

The part of similar weighted value is adjusted in following explanation earlier, because when the lightness of image was very high, the slight change of image was difficult for being distinguished by human eye, so noise is less; Otherwise low when the lightness of pixel, just image is darker, then is easy to generate out bigger noise.Therefore, can use bigger term of reference, make the reference pixel of reconstructed object pixel become many, and then increase the removal anti noise.When the lightness of object pixel is low, use bigger bright deviant (W _Intensity) adjust similar weighted value after, the similar weighted value of gained of tabling look-up is greater can be more higher than the lightness of object pixel the time, make the reference specific gravity of reference pixel improve, and then anti noise is removed in increase, so just can be under the different object pixel of lightness, allow the similar weighted value difference of reference pixel correspondence of identical similarity.

After obtaining bright deviant, the present invention will adjust the similar weighted value (step 220) of each reference pixel of target 101 with bright deviant, in the present embodiment, the mode of adjusting similar weighted value is after adding corresponding bright deviant with the similarity of reference pixel target (S), the method with the similar weighted value of known calculating reference pixel target after adding is produced, that is to say, if known bidirectional filter produces the mode of similar weighted value with RangeTable (S) expression, then present embodiment will be with RangeTable (S+W _Intensity) produce similar weighted value, so the similar weighted value of each reference pixel of object pixel 101 can add drawing with inquiring about of bright deviant by similarity to similar weight table.As shown in Figure 3A, if first reference pixel 102 of object pixel 101 resulting first bright deviant 311 when object pixel 101 is bright is 10, then first reference pixel 102 can use the first bright deviant 311 to adjust the known first similarity weighting curve 310a, makes it to become the second similarity weighting curve 310b as Fig. 3 B; When same object pixel 101 is dark, the second bright deviant 312 that first reference pixel 102 obtains is 30, then first reference pixel 102 can use the second bright deviant 312 to adjust the similar weighting curve 310c that produces after the similarity weight, shown in Fig. 3 C, can learn by the second similar weighting curve 310b and the 3rd weighting curve 310c, when the bright deviant that obtains when first reference pixel 102 is high more, similar weighted value can and then improve, thus, the intensity of similar filter removal picture noise will obtain to promote.

Below the part of distance weighting value is adjusted in explanation, and the present invention can utilize for example known,

(standard deviation) or

Mode such as (gradient), obtain the smoothness of the regional area of rebuilt object pixel, smoothness is used for distinguishing object pixel and is in smooth region or details area, if smooth region, expression can be used near the mode of homogeneous distance weighting and remove noise, makes that level and smooth effect improves in the image, if details area, then keep the effect of using bidirectional filter, fuzzy phenomenon takes place to avoid details area.Because have various images in the image, image and image have the part of boundary, so in an image, certainly will have smooth region and details area simultaneously, because the formed regional area of three reference pixels of corresponding rebuilt object pixel has often comprised smooth region and details area, judge that for fear of using dichotomy object pixel is positioned at smooth region and details area, so after obtaining smoothness, can further use smoothness to the smoothness weight table, to find level and smooth weighted value (W _Smooth), make method that object pixel around smooth region and details area intersection removes picture noise by the filter of homogeneous distance weighting formula excessively to bidirectional filter.

After producing level and smooth weighted value, the present invention can produce the level and smooth weighted value and the calculating formula of the calculating distance weighting value of known bidirectional filter related, makes the distance weighting value of each reference pixel target adjust (step 230) through level and smooth weighted value.In the present embodiment, earlier with formula W _Domain=W _{Domain_prior}+ (256-W _{Domain_prior}) * W _Smooth÷ 256 calculates adjusted distance weighting value, wherein W through the present invention _{Domain_prior}For using the distance weighting value that account form calculated of known bidirectional filter, 256 is the maximum of known distance weighting value.By above-mentioned calculating formula, work as W _SmoothChange into gradually at 256 o'clock by 0, calculate W _DomainThe result that produces of Gaussian filter will change, as shown in Figure 4, the known first distance weighting curve 410a will become second distance weighting curve 410b because of the adjustment of level and smooth weighted value.Because when object pixel is positioned at details area, level and smooth weighted value will be 0, and therefore above-mentioned formula will become W _Domain=W _{Domain_prior}, just will keep using the mode reconstructed object parameter of bidirectional filter this moment, and be positioned at when getting over smooth region when object pixel, smoothly weighted value will be more near 256, and above-mentioned formula makes W _DomainNear 256, work as W more _Domain=256, the distance weighting value of all reference pixels is all identical, the feasible filter of removing the similar average formula of mode of noise.When adjusting distance weighting extremely near the homogeneous weight, can make the reference specific gravity of reference pixel improve, near the effect of using average formula filter, make the noise equalization and reduce, the intensity of removing picture noise apart from filter is increased, and then can revise the problem that known distance weighting value can't effectively reduce the strong noise that high ISO image follows.

After the similar weighted value of all reference pixel target and distance weighting value were all passed through adjustment of the present invention, the present invention can use as known mode reconstructed object pixel (step 240).And object pixel rebuild finish after, select other pixel to rebuild as new object pixel and in above-mentioned mode, up to entire image all rebuild finish after, can produce an image of removing noise, solve the problem that known technology is mentioned.

Have, the method for removal picture noise of the present invention can be implemented in the combination of hardware, software or hardware and software again, also can realize or intersperse among with different assemblies the dispersing mode of the computer system of several interconnected with centralized system and realize in computer system.

Certainly; the present invention also can have other various embodiments; under the situation that does not deviate from spirit of the present invention and essence thereof; those of ordinary skill in the art can make various corresponding changes and distortion according to the present invention, but these corresponding changes and distortion all should belong to the protection range of the appended claim of the present invention.

Claims (3)

1, a kind of method of removing picture noise is characterized in that, comprises the following step:

One of them is an object pixel to comprise in the image of several pixels selected not rebuilt respectively this pixel by one, and sets several reference pixels of this object pixel correspondence;

Calculate this reference pixel respectively respectively to similarity value that should object pixel, convert simultaneously and obtain bright deviant according to the lightness of this object pixel, and with this bright deviant and this similarity value calculating respectively this reference pixel respectively to similar weighted value that should object pixel;

Respectively distance value that should object pixel is calculated this reference pixel respectively respectively to distance weighting value that should object pixel according to this reference pixel respectively;

Calculate a smoothness value of this object pixel, converting according to this smoothness value obtains a level and smooth weighted value, and adjusts this distance weighting value with this level and smooth weighted value; And

Respectively this similar weighted value and adjusted this distance weighting value that should object pixel be rebuild this object pixel with this reference pixel respectively.

2, the method for removal picture noise according to claim 1 is characterized in that, this step of adjusting this distance weighting value also comprises the following step:

One difference of the known maximum of calculating distance weighting value and one of them of described distance weighting value;

Maximum and this difference according to this level and smooth weighted value are calculated an adjustment amount with this level and smooth weighted value; And

This distance weighting value of setting with this distance weighting value and this adjustment amount.

3, the method for removal picture noise according to claim 1, it is characterized in that, this step of rebuilding this object pixel is the summation after the pixel value of this reference pixel respectively and corresponding this distance weighting value and corresponding this level and smooth weighted value are multiplied each other, the summation after multiplying each other divided by corresponding this distance weighting value of this reference pixel respectively and corresponding this level and smooth weighted value.

Images