CN105631825A - Image defogging method based on rolling guidance - Google Patents

Abstract

The invention discloses an image defogging method based on rolling guidance, and mainly solves the problem that the defogging result obtained by using the present defogging method is dark as a whole and accompanied with a vignetting effect. The method is implemented through the following steps of: (1) inputting a foggy color image, and calculating a minimum value image and a dark channel image of the foggy color image in turn; (2) calculating an atmosphere light value of the foggy color image; (3) calculating an initial rough transmission image of the foggy color image; (4) optimizing the initial rough transmission image; (5) calculating an initial defogging result according to an atmospheric scattering physical model; and (6) performing detail enhancement treatment based on layers on the initial defogging result and obtaining a final defogging result. The image defogging method based on rolling guidance avoids the generation of the vignetting effect, improves the whole brightness of the defogging result, makes the defogging result more natural, and can be used in the fields of traffic monitor, automatic drive, space remote sensing and security and protection monitor.

Description

Based on the image defogging method rolling guiding

Technical field

The invention belongs to technical field of image processing, be specifically related to a kind of image defogging method, can be used for traffic monitoring, automatic Pilot, space remote sensing and security monitor field.

Background technology

Extensive persistence haze contamination accident in recent years frequently occurs, everyone healthy and daily life are not only had influence on, and make picture quality serious degradation captured by outdoor monitoring equipment, to outdoor monitoring, monitoring and Intelligent Recognition tracking system etc. bring huge challenge, under affecting in haze weather, the information possibilities such as the car plate of traffic main artery vehicles peccancy cannot accurately identify, the outdoor monitoring equipment of the mechanisms such as bank be likely to because of cannot photograph clearly characteristics of human body's information and cause security protection problem can not effectively obtain guarantee, unmanned plane be likely to because of in air haze interference and can not in taking photo by plane effectively detecting identify the situation such as target. in sum, haze weather processes to outdoor monitoring system later image and introduces a lot of unstable interference factor. therefore, how effectively to remove the interference of haze factor in image and become a problem having important practical significance.

The scholars such as He propose the single image defogging method theoretical based on dark channel prior in article " Singleimagehazeremovalusingdarkchannelprior.IEEETrans.on PatternAnalysisandMachineIntelligence; 2011,33:2341-2353. ". The method is obtained in that very good mist elimination effect, but mist elimination result brightness is partially dark, and owing to adopting soft stingy nomography that initial raw transmission plot is optimized, makes whole defogging method have higher time complexity.

The scholars such as Tarel propose the quick single image defogging method based on twice medium filtering in article " Fastvisibilityrestorationfromasinglecolororgraylevelimag e.ProceedingsofIEEEConferenceonInternationalConferenceon ComputerVision; 2009,10:20-28. ". The method has the mist elimination efficiency that comparison is high, but owing to adopting the median filtering algorithm without good edge retentivity, the marginal area of mist elimination result exists serious halo effect.

Summary of the invention

Present invention aims to the deficiency of above-mentioned conventional images mist elimination technology, it is proposed to a kind of based on the image defogging method rolling guiding, to avoid the generation of halo effect, improve the overall brightness of mist elimination result, the visual effect making mist elimination result is more natural.

For achieving the above object, technical scheme includes as follows:

(1) a width greasy weather coloured image H of input option, calculates the minima image M and dark channel image D of greasy weather coloured image H successively;

(2) the air light value L of greasy weather coloured image H is solved;

(2a) pixel value of pixel in dark channel image D is ranked up by descending, front 0.1% pixel region Q after taking sequence;

(2b) greasy weather coloured image H is transformed in HSV and hue, saturation, intensity space;

(2c) the greasy weather coloured image of position corresponding to the Q of pixel region V-value in HSV space is ranked up by descending, judge the number n being in the greasy weather color image pixel point corresponding to 1/5 position after taking sequence: if n=1, then take this greasy weather color image pixel point as air light value L, if n > 1, then take the maximum greasy weather color image pixel point of pixel value as air light value L;

(3) the initial raw transmission plot of greasy weather coloured image H is solved

$\hat{t}\left(x\right)=\underset{y\∈\mathrm{\Ω}\left(z\right)}{\min }\underset{x\∈\mathrm{\Ω}\left(y\right)}{\max }{t}_{\min }\left(z\right)$

In formula, t_minZ () represents initial raw transmission plotLower limit,Represent y be centered by pixel z, window size be 3 �� 3 local neighborhood in the minimum pixel of pixel value,Represent x be centered by pixel y, window size be the maximum pixel of pixel value in the local neighborhood of 3 �� 3;

(4) to initial raw transmission plotIt is optimized, obtains the transmission plot t (x) after optimizing;

(4a) to initial raw transmission plotCarry out the medium filtering that window size is 15 �� 15, it is thus achieved that filtered transmission plot

(4b) the minima image M according to the following formula, step (1) obtained carries out inversion operation, it is thus achieved that image of the inverted

$\tilde{M}\left(x\right)=255-M\left(x\right),$

In formula, M (x) represents the pixel value of any one pixel x in minima image M,Represent image of the invertedIn the pixel value of any one pixel x;

(4c) with image of the invertedAs input picture, with the transmission plot after medium filteringAs navigational figure, carry out associating bilateral filtering, it is thus achieved that the image after associating bilateral filtering

(4d) with image of the invertedAs input picture, to combine the image after bilateral filteringAs navigational figure, carry out associating bilateral filtering, it is thus achieved that the image after associating bilateral filtering after renewal

(4e) step (4d) is repeated iterative operation, when iteration number of times reaches to set threshold tau, then takes the image after this associating bilateral filteringIt is 5 as transmission plot t (x), the �� value after optimizing;

(5) according to atmosphere light scattering physical model, utilize the transmission plot t (x) after having solved the air light value L obtained and having optimized, obtain initial mist elimination result R₀;

(6) to initial mist elimination result R₀Carry out the details enhancement process based on figure layer, obtain final mist elimination result R.

The present invention compared with prior art has the advantage that

First, the inventive method is optimized by initial raw transmission plot adopts loop iteration associating bilateral filtering, it is to avoid the generation of halo effect, improves the overall brightness of mist elimination result, makes mist elimination result visual effect more natural.

Second, the participation that the method for the present invention not only need not be artificial, moreover it is possible to reduce calculation cost significantly, save the calculating time, while obtaining the visual effect of clear and natural, considerably improve the efficiency of image mist elimination.

Accompanying drawing explanation

Fig. 1 is the flowchart of the present invention;

Fig. 2 is the road traffic Misty Image that the present invention emulates use;

Fig. 3 is the house Misty Image that the present invention emulates use;

Fig. 4 is with the present invention and the existing method mist elimination Comparative result figure to road traffic Misty Image;

Fig. 5 is with the present invention and the existing method mist elimination Comparative result figure to house Misty Image.

Detailed description of the invention

Below in conjunction with accompanying drawing, the present invention will be further described.

With reference to accompanying drawing 1, the present invention specifically comprises the following steps that

Step 1: a width greasy weather coloured image H of input option, calculates the minima image M and dark channel image D of greasy weather coloured image H successively.

Shown in greasy weather the coloured image such as accompanying drawing 2 and accompanying drawing 3 used in the present embodiment. Wherein, accompanying drawing 2 (a) is road traffic Misty Image, is sized to 600 �� 400, and accompanying drawing 3 (a) is house Misty Image, is sized to 441 �� 450.

(1a) from the secondary greasy weather coloured image H as input of the optional width greasy weather coloured image accompanying drawing 2 (a) and accompanying drawing 3 (a);

(1b) each pixel minimum gradation value on R, G, B and three Color Channels of RGB in greasy weather coloured image H is taken, it is thus achieved that the minima image M of greasy weather coloured image H;

(1c) the minima image M of greasy weather coloured image H is carried out the mini-value filtering that window size is 15 �� 15, it is thus achieved that the dark channel image D of greasy weather coloured image H.

Step 2: solve the air light value L of greasy weather coloured image H.

(2a) pixel value of pixel in dark channel image D is ranked up by descending, takes front 0.1% the pixel region Q after sequence;

(2b) greasy weather coloured image H is transformed in HSV and hue, saturation, intensity space;

(2c) the greasy weather coloured image of position corresponding to the Q of pixel region V-value in HSV space is ranked up by descending, judge the number n being in the greasy weather color image pixel point corresponding to 1/5 position after taking sequence: if n=1, then take this greasy weather color image pixel point as air light value L, if n > 1, then take the maximum greasy weather color image pixel point of pixel value as air light value L.

Step 3: solve the initial raw transmission plot of greasy weather coloured image H according to the pixel value edge-restraint condition of pixel in coloured image

$\hat{t}\left(x\right)=\underset{y\∈\mathrm{\Ω}\left(z\right)}{\min }\underset{x\∈\mathrm{\Ω}\left(y\right)}{\max }{t}_{\min }\left(z\right)$

In formula,Represent y be centered by pixel z, window size be 3 �� 3 local neighborhood in the minimum pixel of pixel value,Represent x be centered by pixel y, window size be the maximum pixel of pixel value in the local neighborhood of 3 �� 3, t_minZ () represents initial raw transmission plotLower limit, be calculated as follows:

$t_{\min }\left(z\right)=\min \{\underset{c\∈\{r,g,b\}}{\max }(\frac{L^c-H^c\left(z\right)}{L^c-C_1},\frac{L^c-H^c\left(z\right)}{L^c-C_2}),1\},$

In formula, H (z) represents the pixel value of any one pixel z in greasy weather coloured image H, and { r, g, b} represent that c is any one passage in tri-Color Channels of image R, G, B to c ��, C₁Indicate without the pixel value lower limit of pixel, C in mist coloured image₁Value is 20, C₂Indicate without the pixel value upper limit of pixel, C in mist coloured image₂Value is 280.

Step 4: to initial raw transmission plotIt is optimized, obtains the transmission plot t (x) after optimizing:

(4a) according to the feature that the depth information of coloured image topography block is invariable, to initial raw transmission plotCarry out the medium filtering that window size is 15 �� 15, it is thus achieved that filtered transmission plot

(4b) the minima image M according to the following formula, step 1 obtained carries out inversion operation, it is thus achieved that minima image M image of the inverted

$\tilde{M}\left(x\right)=255-M\left(x\right),$

In formula, M (x) represents the pixel value of any one pixel x in minima image M,Represent image of the invertedIn the pixel value of any one pixel x;

(4c) with minima image M image of the invertedAs input picture, with the transmission plot after medium filteringAs navigational figure, carry out associating bilateral filtering, it is thus achieved that the image after associating bilateral filtering

(4d) with minima image M image of the invertedAs input picture, to combine the image after bilateral filteringAs navigational figure, carry out associating bilateral filtering, it is thus achieved that the image after associating bilateral filtering after renewal

(4e) set threshold tau, iteration optimized after transmission plot t (x):

Due to the transmission plot after medium filteringMarginal information fuzzyyer, according to the associating filter result edge strength that obtains of bilateral filtering feature between input picture and the edge strength of navigational figure, by step (4d) is repeated the transmission plot after iterative operation reinvents medium filteringMarginal information, when iteration number of times reach set threshold tau=5 time, then take this associating bilateral filtering after imageAs the transmission plot t (x) after optimizing.

Step 5: according to atmospheric scattering physical model, obtains initial mist elimination result R₀��

When haze weather, atmospheric scattering physical model is as follows:

H (x)=R₀(x) t (x)+L (1-t (x)),

In formula, H (x) represents the pixel value of any one pixel x in greasy weather coloured image H, and t (x) represents the transmission plot after optimizing, R₀X () represents initial mist elimination result R₀In the pixel value of any one pixel x, L represents air light value.

(5a) conversion of atmospheric scattering physical model is as follows:

$R_0\left(x\right)=\frac{H\left(x\right)-L}{max\left(t\right(x),t_d)}+L,$

In formula, t_dRepresent the restrained boundary of t (x), t_dValue is 0.1;

(5b) step 2 is solved the air light value L obtained and step 4 solve the optimization obtained after transmission plot t (x) substitute into conversion after atmospheric scattering physical model in obtain initial mist elimination result R₀��

Step 6: to initial mist elimination result R₀Carry out the details enhancement process based on figure layer, obtain final mist elimination result R.

(6a) to initial mist elimination result R₀Carry out bilateral filtering, obtain initial mist elimination result R₀Administrative division map layer

(6b) according to the following formula, initial mist elimination result R is obtained₀Detail view layer

$R_0^{\det ail}=R_0-R_0^{smooth};$

(6c) according to the following formula, final mist elimination result R is obtained:

$R=R_0^{smooth}+{\mathrm{\γR}}_0^{\det ail},$

In formula, �� represents that enhancing coefficient, �� value are 1.6.

The effect of the present invention can be described further by following emulation:

1. simulated conditions:

The hardware test platform of the present invention is: processor is InterCorei3350M, and dominant frequency is 2.27GHz, internal memory 2GB, and software platform is: Windows7 Ultimate 32-bit operating system and MatlabR2010b. The input picture of the present invention respectively road traffic Misty Image and house Misty Image, wherein road traffic Misty Image be sized to 600 �� 400, gray level is 256, form is BMP, house Misty Image be sized to 441 �� 450, gray level is 256, and form is BMP.

2. emulation mode:

The single image defogging method based on dark channel prior that the scholars such as method 1:He propose in article " Singleimagehazeremovalusingdarkchannelprior.IEEETrans.on PatternAnalysisandMachineIntelligence; 2011; 33:2341-2353. ", is called for short He method.

The scholars such as method 2:Tarel propose the quick single image defogging method based on twice medium filtering in article " Fastvisibilityrestorationfromasinglecolororgraylevelimag e.ProceedingsofIEEEConferenceonInternationalConferenceon ComputerVision; 2009; 10:20-28. ", are called for short Tarel method.

The single image defogging method based on boundary constraint and contextual information that the scholars such as method 3:Meng propose in article " Efficientimagedehazingwithboundaryconstraintandcontextua lregularization.ComputerVision (ICCV); 2013IEEEInternationalConferenceon.IEEE; 2013:617-624. ", is called for short Meng method.

Method 4: the inventive method.

3. emulation content and interpretation of result:

Experiment 1: by above-mentioned four kinds of methods to carrying out the emulation experiment of image mist elimination for road traffic Misty Image data set shown in Fig. 2, result such as Fig. 4, wherein:

The mist elimination result images that Fig. 4 (a) obtains for He method;

The mist elimination result images that Fig. 4 (b) obtains for Tarel method;

The mist elimination result images that Fig. 4 (c) obtains for Meng method;

The mist elimination result images that Fig. 4 (d) obtains for the inventive method;

As can be seen from Figure 4: the result that the mist elimination algorithm of He et al. obtains, it is fine that color keeps, but is not thoroughly eliminated by interference factor clean; There is halo effect in the result that the mist elimination algorithm of Tarel et al. obtains, causes intensive tree limb and house, distant place still to be disturbed by haze; The result that the mist elimination algorithm of Meng et al. obtains, mist elimination effect is fine, but the vehicle in house at a distance and the middle distant place of road still cannot be clearly visible that; The mist elimination result that the inventive method is obtained, in the middle of road, the vehicle in a distant place, the direction board on highway side and the house of distant place can be clearly visible that.

Experiment 2: by above-mentioned four kinds of methods to carrying out the emulation experiment of image mist elimination for road traffic Misty Image data set shown in Fig. 3, result such as Fig. 5, wherein:

The mist elimination result images that Fig. 5 (a) obtains for He method;

The mist elimination result images that Fig. 5 (b) obtains for Tarel method;

The mist elimination result images that Fig. 5 (c) obtains for Meng method;

The mist elimination result images that Fig. 5 (d) obtains for the inventive method;

As can be seen from Figure 5: the result that the mist elimination algorithm of He et al. obtains, there is higher contrast, but still have mist one layer simple not eliminate; The result that the defogging method of Tarel et al. obtains, still has a lot of fog not remove, and the result integral color obtained has a distortion around leaves; The result that the defogging method of Meng et al. obtains, mist elimination effect is fine, and contrast is higher, but the process of upper right corner houseclearing is excessively bright; The obtained mist elimination result images overall brightness of the inventive method is fine, and color of image also compares nature, and thick grass and leaves marginal information below image are very clear.

Claims (6)

1., based on the image defogging method rolling guiding, comprise the steps:

(1) a width greasy weather coloured image H of input option, calculates the minima image M and dark channel image D of greasy weather coloured image H successively;

(2) the air light value L of greasy weather coloured image H is solved;

(2a) pixel value of pixel in dark channel image D is ranked up by descending, front 0.1% pixel region Q after taking sequence;

(2b) greasy weather coloured image H is transformed in HSV and hue, saturation, intensity space;

(2c) the greasy weather coloured image of position corresponding to the Q of pixel region V-value in HSV space is ranked up by descending, judge the number n being in the greasy weather color image pixel point corresponding to 1/5 position after taking sequence: if n=1, then take this greasy weather color image pixel point as air light value L, if n > 1, then take the maximum greasy weather color image pixel point of pixel value as air light value L;

(3) the initial raw transmission plot of greasy weather coloured image H is solved

$\hat{t}\left(x\right)=\underset{y\∈\mathrm{\Ω}\left(z\right)}{\min }\underset{x\∈\mathrm{\Ω}\left(y\right)}{\max }{t}_{\min }\left(z\right),$

In formula, t_minZ () represents initial raw transmission plotLower limit,Represent y be centered by pixel z, window size be 3 �� 3 local neighborhood in the minimum pixel of pixel value,Represent x be centered by pixel y, window size be the maximum pixel of pixel value in the local neighborhood of 3 �� 3;

(4) to initial raw transmission plotIt is optimized, obtains the transmission plot t (x) after optimizing;

(4a) to initial raw transmission plotCarry out the medium filtering that window size is 15 �� 15, it is thus achieved that filtered transmission plot

(4b) the minima image M according to the following formula, step (1) obtained carries out inversion operation, it is thus achieved that image of the inverted

$\tilde{M}\left(x\right)=255-M\left(x\right),$

In formula, M (x) represents the pixel value of any one pixel x in minima image M,Represent image of the invertedIn the pixel value of any one pixel x;

(4c) with image of the invertedAs input picture, with the transmission plot after medium filteringAs navigational figure, carry out associating bilateral filtering, it is thus achieved that the image after associating bilateral filtering

(4d) with image of the invertedAs input picture, to combine the image after bilateral filteringAs navigational figure, carry out associating bilateral filtering, it is thus achieved that the image after associating bilateral filtering after renewal

(4e) step (4d) is repeated iterative operation, when iteration number of times reaches to set threshold tau=5, then takes the image after this associating bilateral filteringAs the transmission plot t (x) after optimizing;

(5) according to atmospheric scattering physical model, utilize the transmission plot t (x) after having solved the air light value L obtained and having optimized, obtain initial mist elimination result R₀;

(6) to initial mist elimination result R₀Carry out the details enhancement process based on figure layer, obtain final mist elimination result R.

2. according to claim 1 based on the image defogging method rolling guiding, it is characterized in that: step (1) calculates the minima image M of greasy weather coloured image H, it is take each pixel minimum gradation value on R, G, B and three Color Channels of RGB in greasy weather coloured image H, it is thus achieved that the minima image M of greasy weather coloured image H.

3. according to claim 1 based on the image defogging method rolling guiding, it is characterized in that: step (1) calculates the dark channel image D of greasy weather coloured image H, it is that the minima image M to greasy weather coloured image H carries out the mini-value filtering that window size is 15 �� 15, it is thus achieved that the dark channel image D of greasy weather coloured image H.

4. according to claim 1 based on the image defogging method rolling guiding, it is characterised in that: initial raw transmission plot in step (3)Lower limit t_minZ (), is calculated as follows:

$t_{\min }\left(z\right)=min\{\underset{c\∈\{r,g,b\}}{max}(\frac{L^c-H^c\left(z\right)}{L^c-C_1},\frac{L^c-H^c\left(z\right)}{L^c-C_2}),1\},$

In formula, H (z) represents the pixel value of any one pixel z in greasy weather coloured image H, and { r, g, b} represent that c is any one passage in tri-Color Channels of image R, G, B to c ��, C₁Indicate without the pixel value lower limit of pixel, C in mist coloured image₁Value is 20, C₂Indicate without the pixel value upper limit of pixel, C in mist coloured image₂Value is 280.

5. according to claim 1 based on the image defogging method rolling guiding, it is characterised in that: step (5) obtains initial mist elimination result R by atmospheric scattering physical model₀, it is calculated as follows:

$R_0\left(x\right)=\frac{H\left(x\right)-L}{max\left(t\right(x),t_d)}+L,$

In formula, H (x) represents the pixel value of any one pixel x, t in greasy weather coloured image H_dRepresent the restrained boundary of t (x), t_dValue is 0.1, R₀X () represents initial mist elimination result R₀In the pixel value of any one pixel x, L represents air light value.

6. according to claim 1 based on the image defogging method rolling guiding, it is characterised in that: to initial mist elimination result R in step (6)₀Carry out the details enhancement process based on figure layer, carry out as follows:

(6a) to initial mist elimination result R₀Carry out bilateral filtering, obtain initial mist elimination result R₀Administrative division map layer

(6b) according to the following formula, initial mist elimination result R is obtained₀Detail view layer

$R_0^{\det ail}=R_0-R_0^{smooth};$

(6c) according to the following formula, final mist elimination result R is obtained:

$R=R_0^{smooth}+{\mathrm{\γR}}_0^{\det ail},$

In formula, �� represents that enhancing coefficient, �� value are 1.6.