CN104778669A - Fast image denoising method and device - Google Patents

Abstract

The invention provides a fast image denoising method and a fast image denoising device. The fast image denoising method comprises the following steps: calculating a gradient mean of each pixel point of a raw noisy image within a first neighborhood range; judging whether the gradient mean is greater than or equal to a preset first threshold value; if the gradient mean is greater than or equal to the preset first threshold value, respectively calculating a weight between the pixel point and the pixel point in the first neighborhood range, calculating the weighted mean of a gray value of the pixel point, and using the weighted mean of the gray value as an output gray value of the pixel point with the same location of the pixel point in a denoised image; if the gradient mean is less than the preset first threshold value, calculating a gray value mean of the pixel point in a second neighborhood range, and using the gray value mean as the output gray value of the pixel point with the same location of the pixel point in the denoised image. The fast image denoising method disclosed by the invention can filter out the random noise and maintain margins and angular point details of a photographic field in the image.

Description

Rapid image denoising method and device

Technology neighborhood

The present invention relates to image processing techniques neighborhood, particularly relate to a kind of rapid image denoising method and device.

Background technology

Because image capturing system can be subject to the signal disturbing of the various randomness such as temperature, electromagnetic wave, comparatively significantly noise is there is sometimes in the image collected, the many features contained in image can cover by noise, some details in image cannot identification, image visual effect and the quality of data poor.Therefore image processing techniques is studied, weaken random noise to the impact of image, increase contrast and the sharpness of image, ensure image information quality, make computer vision system can also can be reliable, stable under the effect of signal disturbing work, there is very important theory and actual application value undoubtedly.

Although existing image smoothing filtering technique process is simple, but often can only remove extremely strong, that intensity profile is extreme noise, and the process to the smoothing filtering of image, although weaken the impact of noise on the one hand, the edge of original scenery and angle point on the other hand also fuzzy image, lost many detailed information, have impact on visual effect and recognition effect.

Summary of the invention

The invention provides a kind of rapid image denoising method and device, to overcome the problem that prior art easily loses original marginal information in image denoising process.

First aspect, the invention provides a kind of rapid image denoising method, comprising:

Calculate the gradient mean value in the first contiguous range of each pixel of original noisy image, judge whether described gradient mean value is greater than default first threshold; The size of wherein said first neighborhood is preset value;

If described gradient mean value is more than or equal to default first threshold, then calculate the weights of the pixel in described pixel and described first contiguous range respectively, the weights corresponding according to described pixel and gray-scale value calculate the weighted mean value of the gray-scale value of described pixel, using the weighted mean value of described gray-scale value in denoising image with the output gray level value of described pixel at the pixel of same position; If described gradient mean value is less than described default first threshold, then calculate the gray-scale value mean value in the second contiguous range of described pixel, using described gray-scale value mean value in described denoising image with the output gray level value of described pixel at the pixel of same position, the size of wherein said second neighborhood is preset value.

Alternatively, before the weights of the described pixel calculated respectively in described pixel and described first contiguous range, comprising:

Calculate the absolute difference of the gradient mean value in the first contiguous range of the gradient mean value in described first contiguous range of described pixel and the pixel in described first contiguous range respectively;

Judge whether described absolute difference is more than or equal to default Second Threshold.

Alternatively, also comprise:

If described absolute difference is less than described default Second Threshold, the weights of the pixel in described pixel and described first contiguous range are set to 0.

Alternatively, the gradient mean value in the first contiguous range of each pixel of the original noisy image of described calculating, comprising:

Described gradient mean value is calculated by following formula (1):

$\stackrel{\‾}{G}\left(x\right)=\left(\underset{y\∈N\left(x\right)}{\mathrm{\ΣG}\left(y\right)}\right)/s^2---\left(1\right)$

Wherein, x point is arbitrary pixel of described original noisy image, for the gradient mean value in the first contiguous range of x point, N (x) is centered by x point, size is first neighborhood of s × s, y point is centered by x point, point in described first contiguous range, the gray-scale value of gradient image at y point place that G (y) is described original noisy image; S is preset value.

Alternatively, the absolute difference of the gradient mean value in described the first contiguous range calculating the gradient mean value in described first contiguous range of described pixel and the pixel in described first contiguous range respectively, comprising:

Described absolute difference is calculated by following formula (2):

$\mathrm{\ΔG}(x,y)=|\stackrel{\‾}{G}\left(x\right)-\stackrel{\‾}{G}\left(y\right)|---\left(2\right)$

Wherein, for the gradient mean value in the first contiguous range of x point, for the gradient mean value in the first contiguous range of y point, the absolute difference of the gradient mean value in described first contiguous range that Δ G (x, y) is x point and y point.

Alternatively, the weights of the described pixel calculated respectively in described pixel and described first contiguous range, comprising:

If described absolute difference is more than or equal to described default Second Threshold, then calculated the weights between the pixel in described pixel and described first contiguous range by following formula (3):

$W(x,y)=\frac{1}{2{\mathrm{\σ}}^2}{e}^{{\left|\right|N\left(x\right)-N\left(y\right)\left|\right|}_2}---\left(3\right)$

Wherein, ${\left|\right|N\left(x\right)-N\left(y\right)\left|\right|}_2=\underset{i\∈N\left(x\right),j\∈N\left(y\right)}{\mathrm{\Σ}}{(I\left(i\right)-I\left(j\right))}^2/s^2;$

Wherein, x point is more than or equal to the pixel of described default Second Threshold for absolute difference described in described original noisy image, and y point is centered by x point, the point in described first contiguous range; N (x) is centered by x point, and size is first neighborhood of s × s, and N (y) is centered by y point, and size is first neighborhood of s × s; I point is centered by x point, the pixel in described first contiguous range; J point is centered by y point, the pixel in described first contiguous range; I (i) is the gray-scale value of described original noisy image at i point place; I (j) is the gray-scale value of described original noisy image at j point place; S is preset value.

Alternatively, the span of described σ is 10< σ <15.

Alternatively, the described weights corresponding according to described pixel and gray-scale value calculate the weighted mean value of the gray-scale value of described pixel, comprising:

Calculate the accumulated value of weights corresponding to described pixel, comprising:

Described weights W (x, y) is added to matrix W respectively ₀correspondence position in, calculated by following formula (4) and (5):

W ₀(x)＝W ₀(x)+W(x,y) (4)

W ₀(y)＝W ₀(y)+W(x,y) (5)

Wherein, W ₀x () is described matrix W ₀in the value at x point place, W ₀y () is described matrix W ₀in the value at y point place; Described matrix W ₀for with measure-alike, the initialization value of described original noisy image be entirely 0 matrix;

Calculate the weighted accumulation value of the gray-scale value of described each pixel, comprising:

By the product of weights W (x, y) corresponding for described pixel with the gray-scale value of the pixel in the first contiguous range of described pixel, be added to Matrix C respectively ₀correspondence position in, calculated by following formula (6) and (7):

C ₀(x)＝C ₀(x)+W(x,y)×I(y) (6)

C ₀(y)＝C ₀(y)+W(x,y)×I(x) (7)

Wherein, I (x) is the gray-scale value of described original noisy image at x point place, and I (y) is the gray-scale value of described original noisy image at y point place; C ₀x () is described Matrix C ₀in the value at x point place, C ₀y () is described Matrix C ₀in the value at y point place; C ₀for the matrix that, initialization value identical with described original noisy picture size are 0 entirely;

Calculate the weighted mean value of the normalized gray-scale value of described pixel, comprising:

The weighted mean value of the normalized gray-scale value of described pixel is calculated by following formula (8):

I'(v)＝C ₀(v)/W ₀(v) (8)

Wherein, v point is more than or equal to the pixel of described default Second Threshold for absolute difference described in described original noisy image; C ₀v () is described Matrix C ₀in the weighted accumulation value of the gray-scale value of described pixel, W ₀v () is described matrix W ₀at the accumulated value of weights corresponding to described pixel v; I'(v) for described denoising image is in the output gray level value at described pixel v place.

Alternatively, the gray-scale value mean value in the second contiguous range of the described pixel of described calculating, comprising:

Described gray-scale value mean value is calculated by following formula (9):

$I^{\&prime;}\left(u\right)=\left(\underset{y\&Element;\mathrm{\&Psi;}\left(u\right)}{\mathrm{\&Sigma;}}I\left(y\right)\right)/a^2---\left(9\right)$

Wherein, u point is less than the pixel of described default first threshold for gradient mean value described in described original noisy image; I'(u) for described denoising image is at the average gray at u point place; Ψ (u) is centered by u, and size is second neighborhood of a × a; Y point is centered by u point, the point in described second contiguous range; I (y) is the gray-scale value of described original noisy image at y point place; A is preset value.

Alternatively, the span of described s is 15<s<31; The span of described a is: s+5<a<s+10.

Alternatively, described default first threshold is 50.

Second aspect, the invention provides a kind of rapid image denoising device, comprising:

First processing module, for calculating the gradient mean value in the first contiguous range of each pixel of original noisy image, judges whether described gradient mean value is greater than default first threshold; The size of wherein said first neighborhood is preset value;

Second processing module, if be more than or equal to default first threshold for described gradient mean value, then calculate the weights of the pixel in described pixel and described first contiguous range respectively, the weights corresponding according to described pixel and gray-scale value calculate the weighted mean value of the gray-scale value of described pixel, using the weighted mean value of described gray-scale value in denoising image with the output gray level value of described pixel at the pixel of same position; 3rd processing module, if be less than described default first threshold for described gradient mean value, then calculate the gray-scale value mean value in the second contiguous range of described pixel, using described gray-scale value mean value in described denoising image with the output gray level value of described pixel at the pixel of same position, the size of wherein said second neighborhood is preset value.

Rapid image denoising method of the present invention and device, by calculating the gradient mean value in the first contiguous range of each pixel of original noisy image, judge whether described gradient mean value is greater than default first threshold, the size of wherein said first neighborhood is preset value, if described gradient mean value is more than or equal to default first threshold, then calculate the weights of the pixel in described pixel and described first contiguous range respectively, the weights corresponding according to described pixel and gray-scale value calculate the weighted mean value of the gray-scale value of described pixel, using the weighted mean value of described gray-scale value in denoising image with the output gray level value of described pixel at the pixel of same position, if described gradient mean value is less than described default first threshold, then calculate the gray-scale value mean value in the second contiguous range of described pixel, using described gray-scale value mean value in described denoising image with the output gray level value of described pixel at the pixel of same position, the size of wherein said second neighborhood is preset value, can the comparatively serious and random noise of Unknown Distribution of filtering, and while filtering noise, keep edge and the angle point details of scenery in image well, efficiency is higher, solve the problem of easily losing original marginal information in image denoising process.

Accompanying drawing explanation

In order to be illustrated more clearly in the embodiment of the present invention or technical scheme of the prior art, be briefly described to the accompanying drawing used required in embodiment or description of the prior art below, apparently, accompanying drawing in the following describes is some embodiments of the present invention, for this neighborhood those of ordinary skill, under the prerequisite not paying creative work, other accompanying drawing can also be obtained according to these accompanying drawings.

Fig. 1 is the process flow diagram of rapid image denoising method embodiment of the present invention;

Fig. 2 A is the schematic diagram of the original noisy image of the inventive method embodiment;

Fig. 2 B is the schematic diagram of Fig. 2 A after the denoising of existing adaptive median filter method;

Fig. 2 C is the schematic diagram of Fig. 2 A after rapid image denoising method of the present invention denoising;

Fig. 3 A is the schematic diagram of the original noisy image graph of the inventive method embodiment;

Fig. 3 B is the schematic diagram figure of Fig. 3 A after the denoising of existing adaptive median filter method;

Fig. 3 C is the schematic diagram of Fig. 3 A after rapid image denoising method of the present invention denoising;

Fig. 4 A is the schematic diagram of the original noisy image of the inventive method embodiment;

Fig. 4 B is the schematic diagram of Fig. 4 A after the denoising of existing adaptive median filter method;

Fig. 4 C is the schematic diagram of Fig. 4 A after rapid image denoising method of the present invention denoising;

Fig. 5 is the structural representation of rapid image denoising device embodiment of the present invention.

Embodiment

For making the object of the embodiment of the present invention, technical scheme and advantage clearly, below in conjunction with the accompanying drawing in the embodiment of the present invention, technical scheme in the embodiment of the present invention is clearly and completely described, obviously, described embodiment is the present invention's part embodiment, instead of whole embodiments.Based on the embodiment in the present invention, this neighborhood those of ordinary skill, not making the every other embodiment obtained under creative work prerequisite, all belongs to the scope of protection of the invention.

In addition, if below in described each embodiment of the present invention involved technical characteristic do not form conflict each other and just can mutually combine.

Fig. 1 is the process flow diagram of rapid image denoising method embodiment of the present invention, and as shown in Figure 1, the method for the present embodiment can comprise:

Step 101, calculate original noisy image each pixel the first contiguous range in gradient mean value, judge whether described gradient mean value is greater than default first threshold; The size of wherein said first neighborhood is preset value;

If step 102 described gradient mean value is more than or equal to default first threshold, then calculate the weights of the pixel in described pixel and described first contiguous range respectively, the weights corresponding according to described pixel and gray-scale value calculate the weighted mean value of the gray-scale value of described pixel, using the weighted mean value of described gray-scale value in denoising image with the output gray level value of described pixel at the pixel of same position;

If step 103 described gradient mean value is less than described default first threshold, then calculate the gray-scale value mean value in the second contiguous range of described pixel, using described gray-scale value mean value in described denoising image with the output gray level value of described pixel at the pixel of same position, the size of wherein said second neighborhood is preset value.

Thought due to non-local mean filtering can weigh the gray-level structure similarity in region in image well, the weights of pixel are decided by similarity, thus regulate the contribution of different pixels point, and when keeping former graph structure edge and angle point, filter impurity point and noise.But this measurement mode calculated amount is comparatively large, consuming time longer.Therefore, consuming time in order to reduce algorithm, first the complexity in the region of the first contiguous range residing for pixel is simply weighed in the embodiment of the present invention, if the gradient in region is less, can think that pixel is in the region of a gray scale relatively flat, do not need to use the non-local mean filtering algorithm with edge hold facility, only need to use mean filter.

Specifically, the first step calculates the gradient mean value in the first contiguous range residing for each pixel of original noisy image, and the first threshold that setting is preset compares;

If gradient is comparatively large, use non-local mean filtering, even described gradient mean value is more than or equal to default first threshold, then calculate the weights of the pixel in described pixel and described first contiguous range respectively, the weights corresponding according to described pixel and gray-scale value calculate the weighted mean value of the gray-scale value of described each pixel, using the weighted mean value of described gray-scale value in denoising image with the output gray level value of described pixel at the pixel of same position;

If the less use mean filter of gradient, even described gradient mean value is less than default first threshold, then calculate the gray-scale value mean value in the second contiguous range of described pixel, using described gray-scale value mean value in described denoising image with the output gray level value of described pixel at the pixel of same position.

Alternatively, before the weights of the described pixel calculated respectively in described pixel and described first contiguous range, comprising:

Calculate the absolute difference of the gradient mean value in the first contiguous range of the gradient mean value in described first contiguous range of described pixel and the pixel in described first contiguous range respectively;

Judge whether described absolute difference is more than or equal to default Second Threshold.

Alternatively, also comprise:

If described absolute difference is less than described default Second Threshold, the weights of the pixel in described pixel and described first contiguous range are set to 0.

Specifically, if described absolute difference is more than or equal to default Second Threshold, then calculate the weights of the pixel in described pixel and described first contiguous range respectively, the weights corresponding according to described pixel and gray-scale value calculate the weighted mean value of the gray-scale value of described each pixel, using the weighted mean value of described gray-scale value in denoising image with the output gray level value of described pixel at the pixel of same position;

If described absolute difference is less than default Second Threshold, then the weights of the pixel in described pixel and described first contiguous range is set to 0, namely skips the weights accumulation calculating between 2, weights are considered as 0;

Fig. 2 A is the schematic diagram of the original noisy image of the inventive method embodiment, Fig. 2 B is the schematic diagram of Fig. 2 A after the denoising of existing adaptive median filter method, Fig. 2 C is the schematic diagram of Fig. 2 A after rapid image denoising method of the present invention denoising, Fig. 3 A is the schematic diagram of the original noisy image graph of the inventive method embodiment, Fig. 3 B is the schematic diagram figure of Fig. 3 A after the denoising of existing adaptive median filter method, Fig. 3 C is the schematic diagram of Fig. 3 A after rapid image denoising method of the present invention denoising, 4A is the schematic diagram of the original noisy image of the inventive method embodiment, Fig. 4 B is the schematic diagram of Fig. 4 A after the denoising of existing adaptive median filter method, Fig. 4 C is the schematic diagram of Fig. 4 A after rapid image denoising method of the present invention denoising.

As shown in Fig. 2 A, Fig. 2 B, Fig. 2 C, Fig. 2 A is the portrait containing serious chromatic noise after a gray processing, and pixel resolution is 256 × 256; Fig. 2 B is the result images of Fig. 2 A through the denoising of existing adaptive median filter method, calculates consuming timely to be approximately 3.8s; Fig. 2 C is the result images of Fig. 2 A through the denoising of embodiment of the present invention method, calculates about 320ms consuming time.

As shown in Fig. 3 A, Fig. 3 B, Fig. 3 C, Fig. 3 A is the snow scenes figure containing chromatic noise after a gray processing, and pixel resolution is 354 × 221; Fig. 3 B is the result images of Fig. 3 A through the denoising of adaptive median filter method, calculates consuming timely to be approximately 2.5s; Fig. 3 C is the image of Fig. 3 A through the denoising of embodiment of the present invention method, calculates about 560ms consuming time.

As shown in Fig. 4 A, Fig. 4 B, Fig. 4 C, Fig. 4 A is the color handwork picture of the complexity containing chromatic noise after a gray processing, and pixel resolution is 305 × 400, and the details in image has been difficult to because noise effect is serious differentiate.Fig. 4 B is the result images of Fig. 4 A through the denoising of adaptive median filter method, calculates about 7.9s consuming time; Fig. 4 C is the result images of Fig. 4 A through the denoising of embodiment of the present invention method, calculates about 780ms consuming time.

Comparison diagram 2B, Fig. 3 B, Fig. 4 B and Fig. 2 C, 3C, 4C can see, existing adaptive median filter result has remained the impact of some noises, and has factitious piece of phenomenon; And the method denoising result of the embodiment of the present invention from however level and smooth, details is clear, and speed is far away faster than adaptive median filter method, can reach the demand of real-time, interactive denoising.

This transmission embodiment, based on the thought of non-local mean filtering, propose a kind of method of picture noise of effective filtering randomness, can the comparatively serious and random noise of Unknown Distribution of filtering, solve image capturing system and transmission system be subject to randomness, the problem of the noise of Unknown Distribution.

While filtering noise, edge and the angle point details of scenery in image can be kept well, solve the problem of easily losing original marginal information in image denoising process.

The inventive method travelling speed is fast, the successively mathematic calculation such as prescreen and symmetry is utilized to carry out sufficient optimization and acceleration, for colored SD image (720 × 576), can within parallel processing can reach 1s on GPU, and 300ms is only needed for black white image, the needs of real-time, interactive can be met completely, solve non-local mean filtering speed extremely slow, consuming time reaching tens seconds and cause cannot the problem of widespread use.

Alternatively, the gradient mean value in the first contiguous range of each pixel of the original noisy image of described calculating, comprising:

Described gradient mean value is calculated by following formula (1):

$\stackrel{\&OverBar;}{G}\left(x\right)=\left(\underset{y\&Element;N\left(x\right)}{\mathrm{\&Sigma;G}\left(y\right)}\right)/s^2---\left(1\right)$

Wherein, x point is arbitrary pixel of described original noisy image, for the gradient mean value in the first contiguous range of x point, N (x) is centered by x point, size is first neighborhood of s × s, y point is centered by x point, point in described first contiguous range, the gray-scale value of gradient image at y point place that G (y) is described original noisy image; S is preset value.

Specifically, above-mentioned formula (1) is adopted to calculate for the gradient mean value in step 101, namely the gradient mean value in first contiguous range of arbitrary pixel x of described original noisy image is calculated, described first neighborhood is centered by x point, and size is the local rectangular portions of s × s;

Namely final gradient mean value is

Alternatively, the absolute difference of the gradient mean value in described the first contiguous range calculating the gradient mean value in described first contiguous range of described pixel and the pixel in described first contiguous range respectively, comprising:

Described absolute difference is calculated by following formula (2):

$\mathrm{\&Delta;G}(x,y)=|\stackrel{\&OverBar;}{G}\left(x\right)-\stackrel{\&OverBar;}{G}\left(y\right)|---\left(2\right)$

Wherein, for the gradient mean value in the first contiguous range of x point, for the gradient mean value in the first contiguous range of y point, the absolute difference of the gradient mean value in described first contiguous range that Δ G (x, y) is x point and y point.

Specifically, for being judged as the pixel that region gradient is larger in above-mentioned steps 101, namely gradient mean value is more than or equal to default first threshold, and residing region exists the structural informations such as stronger edge and angle point, and use mean filter can these details excessively fuzzy.Therefore these points use non-local mean filtering, and the region that smooth grey is smooth under the condition keeping edge and angle point structure, reaches the object removing noise.

For the pixel using non-local mean filtering, still very large calculated amount is had with each some calculating weights of full figure and intensity-weighted mean value, when exceeding minimizing smooth effect, we can reduce computing time by reducing the point participating in weights contribution.Have the point of many participation weight computing, differ greatly with the gray-level structure in region residing for current pixel point, the weights calculated are very little, contribute negligible.Whether negligible the embodiment of the present invention, by the region gradient size of the first contiguous range of comparison two pixels, prejudge weights.Therefore set the Second Threshold preset of the absolute difference of gradient mean value, the difference between region gradient is less than the point of default Second Threshold, does not participate in weight computing, and the point being more than or equal to default Second Threshold participates in weight computing.

Calculate the absolute difference of the gradient mean value in the first contiguous range of the gradient mean value in described first contiguous range of described pixel and the pixel in described first contiguous range respectively, pixel x is herein the pixel that gradient mean value in the first contiguous range is more than or equal to default first threshold, namely the absolute difference of the gradient mean value in described first contiguous range of x point and y point is calculated by above-mentioned formula (2) with calculate by above-mentioned formula (1) and get.

Alternatively, described default first threshold is 50.

The first threshold preset is 50 is the optimum empirical values be determined by experiment, and can tackle random noise situations in universality ground.

Alternatively, the weights of the described pixel calculated respectively in described pixel and described first contiguous range, comprising:

If described absolute difference is more than or equal to default Second Threshold, then calculated the weights between the pixel in described pixel and described first contiguous range by following formula (3):

$W(x,y)=\frac{1}{2{\mathrm{\&sigma;}}^2}{e}^{{\left|\right|N\left(x\right)-N\left(y\right)\left|\right|}_2}---\left(3\right)$

Wherein, ${\left|\right|N\left(x\right)-N\left(y\right)\left|\right|}_2=\underset{i\&Element;N\left(x\right),j\&Element;N\left(y\right)}{\mathrm{\&Sigma;}}{(I\left(i\right)-I\left(j\right))}^2/s^2;$

Wherein, x point is more than or equal to the pixel of described default Second Threshold for absolute difference described in described original noisy image, and y point is centered by x point, the point in described first contiguous range; N (x) is centered by x point, and size is first neighborhood of s × s, and N (y) is centered by y point, and size is first neighborhood of s × s; I point is centered by x point, the pixel in described first contiguous range; J point is centered by y point, the pixel in described first contiguous range; I (i) is the gray-scale value of described original noisy image at i point place; I (j) is the gray-scale value of described original noisy image at j point place; S is preset value.

Alternatively, the span of described σ is 10< σ <15.

Specifically, for the point participating in weight computing, use Gauss's distance of the area grayscale matrix in two pixel place first contiguous range to weigh, Gauss's distance is calculated as follows:

${\left|\right|N\left(x\right)-N\left(y\right)\left|\right|}_2=\underset{i\&Element;N\left(x\right),j\&Element;N\left(y\right)}{\mathrm{\&Sigma;}}{(I\left(i\right)-I\left(j\right))}^2/s^2$

Gauss is close apart from the gray-level structure in region residing for little explanation two pixels, therefore gives larger weights, otherwise then gives less weights, thus keep original edge and angle point details in gray-level structure.

The weights of the pixel in described pixel and described first contiguous range can be calculated by above-mentioned formula (3).Pixel x is herein the pixel that described absolute difference is more than or equal to default Second Threshold.

σ is smoothness controling parameters, is the optimum empirical value obtained by experiment, can tackle random noise situations in universality ground.

Alternatively, the described weights corresponding according to described pixel and gray-scale value calculate the weighted mean value of the gray-scale value of described pixel, comprising:

Calculate the accumulated value of weights corresponding to described pixel, comprising:

Described weights W (x, y) is added to matrix W respectively ₀correspondence position in, calculated by following formula (4) and (5):

W ₀(x)＝W ₀(x)+W(x,y) (4)

W ₀(y)＝W ₀(y)+W(x,y) (5)

Wherein, W ₀x () is described matrix W ₀in the value at x point place, W ₀y () is described matrix W ₀in the value at y point place; Described matrix W ₀for with measure-alike, the initialization value of described original noisy image be entirely 0 matrix;

Calculate the weighted accumulation value of the gray-scale value of described each pixel, comprising:

By the product of weights W (x, y) corresponding for described pixel with the gray-scale value of the pixel in the first contiguous range of described pixel, be added to Matrix C respectively ₀correspondence position in, calculated by following formula (6) and (7):

C ₀(x)＝C ₀(x)+W(x,y)×I(y) (6)

C ₀(y)＝C ₀(y)+W(x,y)×I(x) (7)

Wherein, I (x) is the gray-scale value of described original noisy image at x point place, and I (y) is the gray-scale value of described original noisy image at y point place; C ₀x () is described Matrix C ₀in the value at x point place, C ₀y () is described Matrix C ₀in the value at y point place; C ₀for the matrix that, initialization value identical with described original noisy picture size are 0 entirely;

Calculate the weighted mean value of the normalized gray-scale value of described pixel, comprising:

The weighted mean value of the normalized gray-scale value of described pixel is calculated by following formula (8):

I'(v)＝C ₀(v)/W ₀(v) (8)

Wherein, v point is more than or equal to the pixel of described default Second Threshold for absolute difference described in described original noisy image; C ₀v () is described Matrix C ₀in the weighted accumulation value of the gray-scale value of described pixel, W ₀v () is described matrix W ₀at the accumulated value of weights corresponding to described pixel v; I'(v) for described denoising image is in the output gray level value at described pixel v place.

Specifically, in order to calculate the weighted mean value of the normalized gray-scale value of described pixel, first the accumulated value of weights corresponding to described pixel is calculated, calculated by above-mentioned formula (4) and (5), in computation process, to each pixel x in image, pixel y in another first contiguous range all to be calculated to the contribution of x.Here described x point and y point are all changes, are the processes of a scanning.X point can scan each pixel of whole image, and this one deck of y point to each given x point, can scan first neighborhood at this current x point place.X and y is the temporary transient code name provided according to the level of scan cycle.

In the computation process of this filtering, be have symmetric.For example, such as we are just at the filter result of calculation level p1=(25,30), so need to participate in a little in the neighborhood of p1 place first calculating, and comprise one of them some p2=(30,35).Calculated p2 to after the contribution of p1, when waiting the filter result of we calculation level p2, p1=(25,30) is in first neighborhood at p2 point place conversely too, so also will calculate a p1 point to the contribution of p2 in order to the result calculating p2 point.

So in fact, have a large amount of calculating repeated, in the mutual contribution of p1 and p2, weights W (p1, p2) equals W's (p2, p1).The embodiment of the present invention is optimized for this point:

When we calculate p1 point filter result, after W (p1, p2) first time is calculated, at once just accordingly its impact in p2 point filter result to adding up into.That is, we are when the filter result calculating p1 point, calculate simultaneously and carried out p1 to it can have influence on contribution a little, so just can also do identical operation when calculating pixel x to pixel y in above-mentioned formula (4)-(7).Time we calculate p2 point like this, p1 point need not be calculated again to the contribution of p2 point.

Represent in above-mentioned formula (4)-(7) be to certain circular treatment pixel x time, have a pixel y in the first neighborhood, be at this moment accumulated x to the contribution of y and y to x by two formula, and be recorded in C ₀in matrix.

The weights W (x, y) corresponding by described pixel is added to matrix W respectively ₀correspondence position in, and by the product of the gray-scale value of the pixel in the first contiguous range of weights W (x, y) corresponding for described pixel and described pixel, be added to Matrix C respectively ₀correspondence position in; The weighted mean value of the normalized gray-scale value of described pixel is calculated finally by formula (8):

I'(v)＝C ₀(v)/W ₀(v)

V point is more than or equal to the pixel of described default Second Threshold for absolute difference described in described original noisy image herein.

If described absolute difference is less than described default Second Threshold, the weights of the pixel in described pixel and described first contiguous range are set to 0.

Utilize the symmetry of above-mentioned weight computing, therefore the embodiment of the present invention adopts the point only calculating half, and scan each pixel successively, the mode of sum weight molecule and denominator is optimized, and decreases only about half of calculated amount.

Alternatively, the gray-scale value mean value in the second contiguous range of the described pixel of described calculating, comprising:

Described gray-scale value mean value is calculated by following formula (9):

$I^{\&prime;}\left(u\right)=\left(\underset{y\&Element;\mathrm{\&Psi;}\left(u\right)}{\mathrm{\&Sigma;}}I\left(y\right)\right)/a^2---\left(9\right)$

Wherein, u point is less than the pixel of described default first threshold for gradient mean value described in described original noisy image; I'(u) for described denoising image is at the average gray at u point place; Ψ (u) is centered by u, and size is second neighborhood of a × a; Y point is centered by u point, the point in described second contiguous range; I (y) is the gray-scale value of described original noisy image at y point place; A is preset value.

Alternatively, the span of described s is 15<s<31; The span of described a is: s+5<a<s+10.

Specifically, for being judged as the pixel that region gradient is less in above-mentioned steps 101, namely described gradient mean value is less than the pixel of described default first threshold, traditional Mean Filtering Algorithm is used just to be enough to smooth noise, use centered by current pixel point, radius is the mean value of the pixel gray-scale value of second neighborhood of a, as the corresponding pixel points of denoising image after filtering noise output gray level value described in average gray can pass through formula (9) calculate get.

Above-mentioned a is a filter radius parameter, and above-mentioned scope is the optimum empirical value be determined by experiment, and can tackle random noise situations in universality ground.

Fig. 5 is the structural representation of rapid image denoising device embodiment of the present invention.As shown in Figure 5, the device of the present embodiment, can comprise: the first processing module 501, second processing module 502 and the 3rd processing module 503; Wherein, described first processing module 501, for calculating the gradient mean value in the first contiguous range of each pixel of original noisy image, judges whether described gradient mean value is greater than default first threshold; The size of wherein said first neighborhood is preset value;

Second processing module 502, if be more than or equal to default first threshold for described gradient mean value, then calculate the weights of the pixel in described pixel and described first contiguous range respectively, the weights corresponding according to described pixel and gray-scale value calculate the weighted mean value of the gray-scale value of described pixel, using the weighted mean value of described gray-scale value in denoising image with the output gray level value of described pixel at the pixel of same position;

3rd processing module 503, if be less than described default first threshold for described gradient mean value, then calculate the gray-scale value mean value in the second contiguous range of described pixel, using described gray-scale value mean value in described denoising image with the output gray level value of described pixel at the pixel of same position, the size of wherein said second neighborhood is preset value.

Alternatively, described second processing module 502, specifically for:

If described gradient mean value is more than or equal to default first threshold, then calculate the absolute difference of the gradient mean value in the first contiguous range of the gradient mean value in described first contiguous range of described pixel and the pixel in described first contiguous range respectively;

Judge whether described absolute difference is more than or equal to default Second Threshold.

Alternatively, described second processing module 502, also for:

If described absolute difference is less than described default Second Threshold, the weights of the pixel in described pixel and described first contiguous range are set to 0.

Alternatively, described first processing module 501, specifically for:

Described gradient mean value is calculated by following formula (1):

$\stackrel{\&OverBar;}{G}\left(x\right)=\left(\underset{y\&Element;N\left(x\right)}{\mathrm{\&Sigma;G}\left(y\right)}\right)/s^2---\left(1\right)$

Wherein, x point is arbitrary pixel of described original noisy image, for the gradient mean value in the first contiguous range of x point, N (x) is centered by x point, size is first neighborhood of s × s, y point is centered by x point, point in described first contiguous range, the gray-scale value of gradient image at y point place that G (y) is described original noisy image; S is preset value.

Alternatively, described second processing module 502, specifically for:

Described absolute difference is calculated by following formula (2):

$\mathrm{\&Delta;G}(x,y)=|\stackrel{\&OverBar;}{G}\left(x\right)-\stackrel{\&OverBar;}{G}\left(y\right)|---\left(2\right)$

Wherein, for the gradient mean value in the first contiguous range of x point, for the gradient mean value in the first contiguous range of y point, the absolute difference of the gradient mean value in described first contiguous range that Δ G (x, y) is x point and y point.

Alternatively, described second processing module 502, specifically for:

If described absolute difference is more than or equal to described default Second Threshold, then calculated the weights between the pixel in described pixel and described first contiguous range by following formula (3):

$W(x,y)=\frac{1}{2{\mathrm{\&sigma;}}^2}{e}^{{\left|\right|N\left(x\right)-N\left(y\right)\left|\right|}_2}---\left(3\right)$

Wherein, ${\left|\right|N\left(x\right)-N\left(y\right)\left|\right|}_2=\underset{i\&Element;N\left(x\right),j\&Element;N\left(y\right)}{\mathrm{\&Sigma;}}{(I\left(i\right)-I\left(j\right))}^2/s^2;$

Wherein, x point is more than or equal to the pixel of described default Second Threshold for absolute difference described in described original noisy image, and y point is centered by x point, the point in described first contiguous range; N (x) is centered by x point, and size is first neighborhood of s × s, and N (y) is centered by y point, and size is first neighborhood of s × s; I point is centered by x point, the pixel in described first contiguous range; J point is centered by y point, the pixel in described first contiguous range; I (i) is the gray-scale value of described original noisy image at i point place; I (j) is the gray-scale value of described original noisy image at j point place; S is preset value.

Alternatively, the span of described σ is 10< σ <15.

Alternatively, described second processing module 502, comprising:

First module, for calculating the accumulated value of weights corresponding to described pixel;

Described first module, specifically for:

Described weights W (x, y) is added to matrix W respectively ₀correspondence position in, calculated by following formula (4) and (5):

W ₀(x)＝W ₀(x)+W(x,y) (4)

W ₀(y)＝W ₀(y)+W(x,y) (5)

Wherein, W ₀x () is described matrix W ₀in the value at x point place, W ₀y () is described matrix W ₀in the value at y point place; Described matrix W ₀for with measure-alike, the initialization value of described original noisy image be entirely 0 matrix;

Second unit, for calculating the weighted accumulation value of the gray-scale value of described each pixel;

Described second unit, specifically for:

By the product of weights W (x, y) corresponding for described pixel with the gray-scale value of the pixel in the first contiguous range of described pixel, be added to Matrix C respectively ₀correspondence position in, calculated by following formula (6) and (7):

C ₀(x)＝C ₀(x)+W(x,y)×I(y) (6)

C ₀(y)＝C ₀(y)+W(x,y)×I(x) (7)

Wherein, I (x) is the gray-scale value of described original noisy image at x point place, and I (y) is the gray-scale value of described original noisy image at y point place; C ₀x () is described Matrix C ₀in the value at x point place, C ₀y () is described Matrix C ₀in the value at y point place; C ₀for the matrix that, initialization value identical with described original noisy picture size are 0 entirely;

Unit the 3rd, for calculating the weighted mean value of the normalized gray-scale value of described pixel;

Described Unit the 3rd, specifically for:

The weighted mean value of the normalized gray-scale value of described pixel is calculated by following formula (8):

I'(v)＝C ₀(v)/W ₀(v) (8)

Wherein, v point is more than or equal to the pixel of described default Second Threshold for absolute difference described in described original noisy image; C ₀v () is described Matrix C ₀in the weighted accumulation value of the gray-scale value of described pixel, W ₀v () is described matrix W ₀at the accumulated value of weights corresponding to described pixel v; I'(v) for described denoising image is in the output gray level value at described pixel v place.

Alternatively, described 3rd processing module, specifically for:

Described gray-scale value mean value is calculated by following formula (9):

$I^{\&prime;}\left(u\right)=\left(\underset{y\&Element;\mathrm{\&Psi;}\left(u\right)}{\mathrm{\&Sigma;}}I\left(y\right)\right)/a^2---\left(9\right)$

Wherein, u point is less than the pixel of described default first threshold for gradient mean value described in described original noisy image; I'(u) for described denoising image is at the average gray at u point place; Ψ (u) is centered by u, and size is second neighborhood of a × a; Y point is centered by u point, the point in described second contiguous range; I (y) is the gray-scale value of described original noisy image at y point place; A is preset value.

Alternatively, the span of described s is 15<s<31; The span of described a is: s+5<a<s+10.

Alternatively, described default first threshold is 50.

The device of the present embodiment, may be used for the technical scheme performing embodiment of the method shown in Fig. 1, it realizes principle and technique effect is similar, repeats no more herein.

This neighborhood those of ordinary skill is understood that all or part of step realizing above-mentioned each embodiment of the method can have been come by the hardware that programmed instruction is relevant.Aforesaid program can be stored in a computer read/write memory medium.This program, when performing, performs the step comprising above-mentioned each embodiment of the method; And aforesaid storage medium comprises: ROM, RAM, magnetic disc or CD etc. various can be program code stored medium.

Last it is noted that above each embodiment is only in order to illustrate technical scheme of the present invention, be not intended to limit; Although with reference to foregoing embodiments to invention has been detailed description, the those of ordinary skill of this neighborhood is to be understood that: it still can be modified to the technical scheme described in foregoing embodiments, or carries out equivalent replacement to wherein some or all of technical characteristic; And these amendments or replacement, do not make the essence of appropriate technical solution depart from the scope of various embodiments of the present invention technical scheme.

Claims (12)

1. a rapid image denoising method, is characterized in that, comprising:

Calculate the gradient mean value in the first contiguous range of each pixel of original noisy image, judge whether described gradient mean value is more than or equal to default first threshold; The size of wherein said first neighborhood is preset value;

If described gradient mean value is more than or equal to default first threshold, then calculate the weights of the pixel in described pixel and described first contiguous range respectively, the weights corresponding according to described pixel and gray-scale value calculate the weighted mean value of the gray-scale value of described pixel, using the weighted mean value of described gray-scale value in denoising image with the output gray level value of described pixel at the pixel of same position;

If described gradient mean value is less than described default first threshold, then calculate the gray-scale value mean value in the second contiguous range of described pixel, using described gray-scale value mean value in described denoising image with the output gray level value of described pixel at the pixel of same position, the size of wherein said second neighborhood is preset value.

2. method according to claim 1, is characterized in that, before the weights of the described pixel calculated respectively in described pixel and described first contiguous range, comprising:

Calculate the absolute difference of the gradient mean value in the first contiguous range of the gradient mean value in described first contiguous range of described pixel and the pixel in described first contiguous range respectively;

Judge whether described absolute difference is more than or equal to default Second Threshold.

3. method according to claim 2, is characterized in that, also comprises:

If described absolute difference is less than described default Second Threshold, the weights of the pixel in described pixel and described first contiguous range are set to 0.

4. the method according to any one of claim 1-3, is characterized in that, the gradient mean value in the first contiguous range of each pixel of the original noisy image of described calculating, comprising:

Described gradient mean value is calculated by following formula (1):

$\stackrel{\&OverBar;}{G}\left(x\right)=\left(\underset{y\&Element;N\left(x\right)}{\mathrm{\&Sigma;G}\left(y\right)}\right)/s^2---\left(1\right)$

Wherein, x point is arbitrary pixel of described original noisy image, for the gradient mean value in the first contiguous range of x point, N (x) is centered by x point, size is first neighborhood of s × s, y point is centered by x point, point in described first contiguous range, the gray-scale value of gradient image at y point place that G (y) is described original noisy image; S is preset value.

5. according to the method in claim 2 or 3, it is characterized in that, the absolute difference of the gradient mean value in described the first contiguous range calculating the gradient mean value in described first contiguous range of described pixel and the pixel in described first contiguous range respectively, comprising:

Described absolute difference is calculated by following formula (2):

$\mathrm{\&Delta;G}(x,y)=|\stackrel{\&OverBar;}{G}\left(x\right)-\stackrel{\&OverBar;}{G}\left(y\right)|---\left(2\right)$

Wherein, for the gradient mean value in the first contiguous range of x point, for the gradient mean value in the first contiguous range of y point, the absolute difference of the gradient mean value in described first contiguous range that Δ G (x, y) is x point and y point.

6. according to the method in claim 2 or 3, it is characterized in that, the weights of the described pixel calculated respectively in described pixel and described first contiguous range, comprising:

If described absolute difference is more than or equal to described default Second Threshold, then calculated the weights between the pixel in described pixel and described first contiguous range by following formula (3):

$W(x,y)=\frac{1}{{2\mathrm{\&sigma;}}^2}{e}^{{\left|\right|N\left(x\right)-N\left(y\right)\left|\right|}_2}---\left(3\right)$

Wherein, ${\left|\right|N\left(x\right)-N\left(y\right)\left|\right|}_2=\underset{i\&Element;N\left(x\right),j\&Element;N\left(y\right)}{\mathrm{\&Sigma;}}{(I\left(i\right)-I\left(j\right))}^2/s^2;$

Wherein, x point is more than or equal to the pixel of described default Second Threshold for absolute difference described in described original noisy image, and y point is centered by x point, the point in described first contiguous range; N (x) is centered by x point, and size is first neighborhood of s × s, and N (y) is centered by y point, and size is first neighborhood of s × s; I point is centered by x point, the pixel in described first contiguous range; J point is centered by y point, the pixel in described first contiguous range; I (i) is the gray-scale value of described original noisy image at i point place; I (j) is the gray-scale value of described original noisy image at j point place; S is preset value.

7. method according to claim 6, is characterized in that, the span of described σ is 10< σ <15.

8. method according to claim 6, is characterized in that, the described weights corresponding according to described pixel and gray-scale value calculate the weighted mean value of the gray-scale value of described pixel, comprising:

Calculate the accumulated value of weights corresponding to described pixel, comprising:

Described weights W (x, y) is added to matrix W respectively ₀correspondence position in, calculated by following formula (4) and (5):

W ₀(x)＝W ₀(x)+W(x,y) (4)

W ₀(y)＝W ₀(y)+W(x,y) (5)

Wherein, W ₀x () is described matrix W ₀in the value at x point place, W ₀y () is described matrix W ₀in the value at y point place; Described matrix W ₀for with measure-alike, the initialization value of described original noisy image be entirely 0 matrix;

Calculate the weighted accumulation value of the gray-scale value of described each pixel, comprising:

By the product of weights W (x, y) corresponding for described pixel with the gray-scale value of the pixel in the first contiguous range of described pixel, be added to Matrix C respectively ₀correspondence position in, calculated by following formula (6) and (7):

C ₀(x)＝C ₀(x)+W(x,y)×I(y) (6)

C ₀(y)＝C ₀(y)+W(x,y)×I(x) (7)

Wherein, I (x) is the gray-scale value of described original noisy image at x point place, and I (y) is the gray-scale value of described original noisy image at y point place; C ₀x () is described Matrix C ₀in the value at x point place, C ₀y () is described Matrix C ₀in the value at y point place; C ₀for the matrix that, initialization value identical with described original noisy picture size are 0 entirely;

Calculate the weighted mean value of the normalized gray-scale value of described pixel, comprising:

The weighted mean value of the normalized gray-scale value of described pixel is calculated by following formula (8):

I'(v)＝C ₀(v)/W ₀(v) (8)

Wherein, v point is more than or equal to the pixel of described default Second Threshold for absolute difference described in described original noisy image; C ₀v () is described Matrix C ₀in the weighted accumulation value of the gray-scale value of described pixel, W ₀v () is described matrix W ₀at the accumulated value of weights corresponding to described pixel v; I'(v) for described denoising image is in the output gray level value at described pixel v place.

9. the method according to any one of claim 1-3, is characterized in that, the gray-scale value mean value in the second contiguous range of the described pixel of described calculating, comprising:

Described gray-scale value mean value is calculated by following formula (9):

$I^{\&prime;}\left(u\right)=\left(\underset{y\&Element;\mathrm{\&Psi;}\left(u\right)}{\mathrm{\&Sigma;}}I\left(y\right)\right)/a^2---\left(9\right)$

Wherein, u point is less than the pixel of described default first threshold for gradient mean value described in described original noisy image; I'(u) for described denoising image is at the average gray at u point place; Ψ (u) is centered by u, and size is second neighborhood of a × a; Y point is centered by u point, the point in described second contiguous range; I (y) is the gray-scale value of described original noisy image at y point place; A is preset value.

10. the method according to any one of claim 1-3, is characterized in that, the span of described s is 15<s<31; The span of described a is: s+5<a<s+10.

11. methods according to any one of claim 1-3, it is characterized in that, described default first threshold is 50.

12. 1 kinds of rapid image denoising devices, is characterized in that, comprising:

First processing module, for calculating the gradient mean value in the first contiguous range of each pixel of original noisy image, judges whether described gradient mean value is greater than default first threshold; The size of wherein said first neighborhood is preset value;

Second processing module, if be more than or equal to default first threshold for described gradient mean value, calculate the weights of the pixel in described pixel and described first contiguous range respectively, the weights corresponding according to described pixel and gray-scale value calculate the weighted mean value of the gray-scale value of described pixel, using the weighted mean value of described gray-scale value in denoising image with the output gray level value of described pixel at the pixel of same position;

3rd processing module, if be less than described default first threshold for described gradient mean value, then calculate the gray-scale value mean value in the second contiguous range of described pixel, using described gray-scale value mean value in described denoising image with the output gray level value of described pixel at the pixel of same position, the size of wherein said second neighborhood is preset value.