CN105913391A - Defogging method based on shape variable morphological reconstruction - Google Patents

Abstract

The invention discloses a defogging method based on shape variable morphological reconstruction, comprising steps of using an image smallest color channel and a biggest color channel while synchronously performing parallel calculation on corresponding medium transmissivity, optimizing transmissivity of the smallest color channel through morphological expansion reconstruction, performing fusion on two transmissivity distribution graphs, and obtaining an optimal transmissivity which is immune to sky. Compared with the current medium transmissivity refinement method in a defogging algorithm, the defogging method of the invention can fundamentally inhibit a halo phenomenon, does not need to detect in advance, adaptively compensates the transmissivity of sky area low estimation and avoids distortion of the sky area in the defogging result.

Description

A kind of defogging method based on the variable Morphological Reconstruction of shape

Technical field

The present invention relates to technical field of image recovery, specifically disclose a kind of defogging method.

Background technology

In recent years, air quality declines serious, and the bad weather such as mist and haze frequently occurs.Being affected by haze weather, image presents low Contrast, low definition feature, at close shot, detailed information is lost serious, and at distant view, feature is completely covered or fuzzy, information can Identification is substantially reduced.Color fidelity declines simultaneously, serious color offset phenomenon occurs.

Image mist elimination is by certain technological means, removes the fog interference in image, in order to obtain being satisfied with visual effect and obtaining more Many effective informations.He et al. proposes dark channel prior statistical law, and proposes " dark channel prior+guidance quality filter on this basis Ripple " algorithm, it is acknowledged as the algorithm of current mist elimination best results.

But, the essence of guidance quality filtering is to be operated by the mean filter of large scale window, and energy concentrates on the " light at edge originally Dizzy " diffusion, range of scatter is the biggest, and the halation intensity of edge is the least, weakens, with this, the halation phenomenon (halo that depth of field sudden change produces Effect), range of scatter size is directly determined by mean filter window size.The adjoint side effect of this operation is: the halation of diffusion is again Obscure the scene depth of adjacent edges, the transmission scene existed particularly with intermittence scenery such as the woods, branches and leaves, trunks, and The edge of depth of field acute variation.Sudden change region intensive for the degree of depth, the halation of diffusion " has filled up " depth of field difference in gap；For deeply Degree acute variation region, the halation of diffusion makes the depth of field smooth variation of precipitous change.The transmissivity that mistake is estimated causes this subregion Can not reach well to remove fog effect.It is interpreted as with Formula Solution: t (x)=e^-βd(x), halation that gap area spreads due to mean filter Making d (x) less than normal, transmissivity is bigger than normal, and generally I (x)-A≤0 at big depth of field sudden change, thereforeThe most inclined Big transmissivity makes this partial pixel value bigger than normal, close to air light value, presents " greyish white vaporific ", does not reaches fog effect.Its Secondary, not to be noted transmissivity is only relevant with scene depth, with image close grain structure without direct relation.Strong structural texture correspondence scene The place that degree of depth break edge, i.e. halation phenomenon occur, and the same degree of depth in the generally corresponding scene of fine texture texture.Therefore fine texture Texture is not only unrelated with transmissivity, also results in transmissivity unsmooth, there is the effect of similar " interference noise ".Finally, to being unsatisfactory for The sky areas of dark channel prior does not deals with, and easily causes sky areas color distortion.

Summary of the invention

It is an object of the invention to provide a kind of defogging method based on the variable Morphological Reconstruction of shape, deposit solving existing mist elimination algorithm Problem.The present invention proposes to utilize deformable structure unit adaptive median filter to combine gray scale morphology expansion restructing algorithm and becomes more meticulous Minimum Color Channel transmissivity, suppresses halation phenomenon；Calculate original has mist image maximum color channel image simultaneously, and estimated brightness is relatively The transmissivity in high region；By the mixing operation to two width transmissivity distribution maps, it is thus achieved that without blocking effect, adaptive equalization sky areas Optimal transmission rate, make mist elimination result clear.

To achieve these goals, the present invention adopts the following technical scheme that

A kind of defogging method based on the variable Morphological Reconstruction of shape, comprises the following steps:

Step 1: have mist image to original, asks the three-channel minimum pixel value of each point R, G, B as minimum color channel image Dc, max pixel value, as maximum color channel image mc, carries out local minimum value filtering to minimum color channel image, it is thus achieved that Rough estimate dark primary channel image DarkImg；

Step 2: determine atmosphere light position candidate from rough estimate dark primary channel image DarkImg, original have in mist image seek Correspondence is looked for have the value of maximum brightness point as air light value A；

Step 3: have the gray-scale map of mist image for guiding figure with original, calculates deformable structure element, with the structural element pair generated Minimum color channel image dc carries out adaptive median filter operation, it is thus achieved that repressed structural images dc_med of details；

Step 4: rough estimate dark primary channel image DarkImg is reconstruct mark image in step 1, with structure chart in step 3 As dc_med is reconstruct template image, carries out gray scale morphology and expand reconstruct, it is thus achieved that dark channel image DarkImg* become more meticulous； In conjunction with air light value A, utilizeCalculate the minimum Color Channel transmissivity distribution map t_dc become more meticulous；

Step 5: maximum color channel image mc is performed computingObtain maximum color passage transmissivity distribution map t_mc； T is threshold value；

Step 6: to the minimum Color Channel transmissivity distribution map t_dc become more meticulous, maximum color passage transmissivity distribution map t_mc Perform point-by-point comparison and take the mixing operation of maximum, it is thus achieved that the optimal transmission rate distribution map t* to sky immunity；

Step 7: directly utilizing air light value A and optimal transmission rate t* has mist image to carry out mist elimination sharpening to original, obtains mist elimination Reconstruct image J (x):

$J\left(x\right)=\frac{I\left(x\right)-A}{t\left(x\right)}+A---\left(7\right).$

Further, carrying out in step 1 arranging window size during local minimum value filtering is 7 pixel units.

Further, step 2 is particularly as follows: take the pixel that 0.1% brightness is maximum from rough estimate dark primary channel image DarkImg Determine atmosphere light position candidate, have searching correspondence in mist image to have the value of maximum brightness point as air light value A original.

Further, in step 3:

d_λRepresenting deformable structure unit length, definition mode is as follows:

$\left\{\begin{array}{c}{d}_{\mathrm{\λ}}(x,x)=0\\ d_{\mathrm{\λ}}(x,y)=\underset{\mathrm{\σ}}{\min }\left\{L\left(\mathrm{\σ}\right)\right\}\end{array}\right.---\left(1\right)$

Wherein, the passage length between L (σ) representative point x and some y:

$L\left(\mathrm{\σ}\right)=\underset{i=0}{\overset{n}{\Σ}}\[1+\mathrm{\λ}\·d_{pixel}\left(\mathrm{Im}age\right(x_i),\mathrm{Im}age(x_{i+1}\left)\right)\]---\left(2\right)$

d_pixelFor gray scale difference value, σ=(x=x0, x1 ... xn=y) represents the passage between some x and some y, and λ is arithmetic number parameter； When λ=0, deformable structure element is the whole filter kernel region of definition, is equivalent to fixed structure element；

Select original have the gray-scale map of mist image as guiding figure；When calculating deformable structure unit, set to deformable structure unit length Put the upper limit；

A_λ,r=y | d_λ(x,y)≤r} (3)

R is structural elements length limit；λ=0.5；

$r=\frac{\underset{y\∈\mathrm{\Ω}\left(x\right)}{\Σ}{d}_{\mathrm{\λ},y}}{N}---\left(4\right)$

Deformable structure unit determines, minimum color channel image dc is carried out medium filtering；Obtain the repressed knot of fine texture texture Composition is as dc_med.

Further, in step 5, threshold value T is 170.

Relative to prior art, the method have the advantages that

(1) boundary alignment template image accurately limits, and at depth of field sudden change, the shallower pixel of the depth of field does not interferes with the bigger picture of the depth of field The dark value of element, fundamentally suppresses transmissivity change at halation phenomenon, and sudden change precipitous.

(2) adaptive median filter based on deformable structure unit, has filtered fine texture texture, makes transmissivity smooth, and mist elimination is tied The grain details of fruit is apparent.

(3), in transmittance calculation, in conjunction with original image maximum color passage distribution map, adaptive equalization sky areas transmissivity, keep away Remove mist result sky areas color distortion from.

Accompanying drawing explanation

Fig. 1 is the flow chart of defogging method of the present invention；

Fig. 2 is the detail flowchart of defogging method of the present invention；

Fig. 3 is embodiment 1 fog-degraded image and each link mist elimination intermediate result image；Wherein, Fig. 3 (a) is original to have mist figure Picture, Fig. 3 (b) is dark primary channel image, and Fig. 3 (c) is minimum Color Channel transmissivity distribution map t_dc, and Fig. 3 (d) is Maximum color passage transmissivity distribution map t_mc, Fig. 3 (e) be to sky immunity optimal transmission rate distribution t, Fig. 3 (f) be Whole mist elimination result images.

Detailed description of the invention

The present invention is elaborated by explanation and detailed description of the invention below in conjunction with the accompanying drawings.

The atmospherical scattering model proposed according to McCartney, inversion chart picture, in the process that degrades in greasy weather, obtains scene imageIt is satisfied by [0,255] pixel boundary according to each passage of image J to limit, obtains transmissivity inequality as follows:

$\underset{c\∈\{r,g,b\}}{min}\underset{p\∈B}{min}\left(J_c\right(p\left)\right)\≥0\⇒\underset{c\∈\{r,g,b\}}{min}\underset{p\∈B}{min}\left(\frac{1}{t}\right(I\left(p\right)-A_c)+A_c)\≥0\⇒t\≥\underset{c\∈\{r,g,b\}}{min}\underset{p\∈B}{min}\left\{\frac{I\left(p\right)-A_c}{-A_c}\right\}$

$\underset{c\∈\{r,g,b\}}{\max }\underset{p\∈B}{\max }\left(J_c\right(p\left)\right)\≤255\⇒\underset{c\∈\{r,g,b\}}{\max }\underset{p\∈B}{\max }\left(\frac{1}{t}\right(I\left(p\right)-A_c)+A_c)\≤255\⇒t\≥\underset{c\∈\{r,g,b\}}{\max }\underset{p\∈B}{\max }\left\{\frac{I\left(p\right)-A_c}{255-A_c}\right\}$

The local contrast strengthened in conjunction with mist elimination is inversely proportional to transmissivity, and obtaining optimal transmission rate is:

$t*=max\left\{\underset{c\∈\{r,g,b\}}{min}\underset{p\∈B}{min}\right\{\frac{I\left(p\right)-A_c}{-A_c}\},\underset{c\∈\{r,g,b\}}{max}\underset{p\∈B}{max}\{\frac{I\left(p\right)-A_c}{255-A_c}\left\}\right\}$

In formula, Section 1 is the transmissivity that minimum color channel image is corresponding, and Section 2 is the transmissivity that maximum color passage is corresponding. I(p)-A_cThe purpose of operation can be regarded as obtaining the sky areas of high brightness, and Ac is threshold value.But directly with atmosphere light Value is as threshold value, it is thus achieved that high-brightness region limited, need suitably to reduce threshold value, herein temporarily using T as threshold value, Air light value in replacement formula.

When performing defogging, it usually needs transmissivity arranges lower limit, experiment finds t₀=0.25 is obtained in that preferable mist elimination Result.I.e. t₀=0.25 is the minimum of a value of minimum Color Channel transmissivity, can calculate the maximum gradation value that image is corresponding:

$\frac{I\left(p\right)-255}{-255}\≥0.25\⇒I\left(p\right)\≤191$

Meanwhile, t₀=0.25 is the minimum of a value that maximum color passage obtains transmissivity, in conjunction with I (p)≤191, takes the situation of I (p)=191:

$\frac{191-T}{255-T}\≥0.25\⇒T\≤170$

Therefore, threshold T=170 are taken.Owing to image maximum color passage purpose is the transmissivity calculating brightness value upper zone, should The feature in region is that gray value is higher and change is mild, therefore can omit local maximum filtering operation, directly utilize maximum color Path computation transmissivity.

The optimal transmission rate of final simplification is:

$t*=max\left\{\underset{c\∈\{r,g,b\}}{min}\underset{p\∈B}{min}\right\{\frac{I_c\left(p\right)-255}{-255}\},\underset{c\∈\{r,g,b\}}{max}\{\frac{I_c-170}{255-170}\left\}\right\}$

Image border change is included in as connectivity criteria by deformable structure unit, in conjunction with spatial domain distance and pixel value distance, instructs raw Become to adapt to the structural element of image local texture.Adaptive median filter based on deformable structure unit, it is possible at suppression close grain Retain strong edge while structure not to be attenuated, and edge accurate positioning, it is achieved protect limit noise removal function.

A kind of defogging method based on the variable Morphological Reconstruction of shape of the present invention, comprises the steps:

Step 1: have mist image to original, asks the three-channel minimum pixel value of each point R, G, B as minimum color channel image Dc, max pixel value, as maximum color channel image mc, carries out local minimum value filtering to minimum color channel image, it is thus achieved that Rough estimate dark primary channel image DarkImg；

Step 2: utilize He atmosphere light to obtain way (Kaiming He, Jian Sun, Xiaoou Tang, " Single Image Haze Removal Using Dark Channel Prior ", 2009CVPR), from rough estimate dark primary channel image DarkImg, take 0.1% The pixel of brightness maximum determines atmosphere light position candidate, has searching correspondence in mist image to have the value of maximum brightness point as greatly original Gas light value A；

Step 3: have the gray-scale map of mist image for guiding figure with original, calculates deformable structure element, with the structural element pair generated Minimum color channel image dc carries out adaptive median filter operation, it is thus achieved that repressed structural images dc_med of details；

Step 4: rough estimate dark primary channel image DarkImg is reconstruct mark image in step 1, with structure chart in step 3 As dc_med is reconstruct template image, carries out gray scale morphology and expand reconstruct, it is thus achieved that dark channel image DarkImg* become more meticulous； In conjunction with air light value A, utilizeCalculate the minimum Color Channel transmissivity distribution map t_dc become more meticulous；

Step 5: maximum color channel image mc is performed computingObtain maximum color passage transmissivity distribution map t_mc；

Step 6: to the minimum Color Channel transmissivity distribution map t_dc become more meticulous, maximum color passage transmissivity distribution map t_mc Perform point-by-point comparison and take the mixing operation of maximum, it is thus achieved that the optimal transmission rate distribution map t* to sky immunity；The minimum of this transmissivity Value is 0.25, performs defogging without again transmissivity being arranged lower limit；

Step 7: without again transmissivity being arranged lower limit, directly utilizing air light value A and optimal transmission rate t* has mist image to original Carry out mist elimination sharpening；

$J\left(x\right)=\frac{I\left(x\right)-A}{t\left(x\right)}+A$

Application the inventive method, carries out mist elimination process to Fig. 3 (a), specifically comprises the following steps that

Step 1, calculates dark primary channel image DarkImg.There is mist image I to original, compare each point R, G, B three-channel Minimum pixel value is as minimum Color Channel dc, and the three-channel max pixel value of each point R, G, B is as maximum color passage mc. Dc image is performed local minimum filtering operation, it is thus achieved that dark primary channel image DarkImg, arrange window size is 7 herein Individual pixel unit, shown in result such as Fig. 3 (b)；

Step 2, determines air light value A.The pixel that 0.1% brightness is maximum is taken from rough estimate dark primary channel image DarkImg Determine atmosphere light position candidate, have searching correspondence in mist image to have the value of maximum brightness point as air light value A original；

Step 3, calculates deformable structure element, performs adaptive median filter, it is thus achieved that repressed structural images dc_med of details. d_λRepresenting deformable structure unit length, definition mode is as follows:

$\left\{\begin{array}{c}{d}_{\mathrm{\λ}}(x,x)=0\\ d_{\mathrm{\λ}}(x,y)=\underset{\mathrm{\σ}}{\min }\left\{L\left(\mathrm{\σ}\right)\right\}\end{array}\right.---\left(1\right)$

Wherein, the passage length between L (σ) representative point x and some y:

$L\left(\mathrm{\σ}\right)=\underset{i=0}{\overset{n}{\Σ}}\[1+\mathrm{\λ}\·d_{pixel}\left(\mathrm{Im}age\right(x_i),\mathrm{Im}age(x_{i+1}\left)\right)\]---\left(2\right)$

d_pixelFor gray scale difference value, σ=(x=x0, x1 ... xn=y) represents the passage between some x and some y, and λ is arithmetic number parameter.When During λ=0, deformable structure element is the whole filter kernel region of definition, is equivalent to fixed structure element.Original noisy image, Owing to noise can affect the growth of structural elements, needing to calculate structural elements shape in another piece image, this image is called guiding figure.This Place, selects original have the gray-scale map of mist image as guiding figure.When calculating deformable structure unit, need to deformable structure unit head Degree arranges the upper limit, because may become very big in flat site structural elements, causes calculating time length computationally intensive.

A_λ,r=y | d_λ(x,y)≤r} (3)

R is structural elements length limit.In the rectangular extent herein selecting the length of side to be 5, the average of each point structural elements length is as length limit, Gradient weight λ=0.5 is set.

$r=\frac{\underset{y\∈\mathrm{\Ω}\left(x\right)}{\Σ}{d}_{\mathrm{\λ},y}}{N}---\left(4\right)$

Deformable structure unit determines, minimum color channel image dc is carried out medium filtering.Obtain the repressed knot of fine texture texture Composition is as dc_med.

Step 4, morphological dilations reconstructs the dark figure that becomes more meticulous, and calculates minimum Color Channel transmissivity distribution.With thick in step 1 Estimating that dark primary channel image DarkImg marks image for reconstruct, in step 3, structural images dc_med is reconstruct template image, Carry out gray scale morphology and expand reconstruct, it is thus achieved that dark channel image DarkImg* become more meticulous；In conjunction with air light value A, calculate fine Shown in the minimum Color Channel transmissivity distribution map t_dc, result such as Fig. 3 (c) changed.

$t\_dc=1-\frac{Dark\mathrm{Im}g*}{A}---\left(5\right)$

Step 5, calculates the distribution of maximum color passage transmissivity.Step 1 maximum color channel image mc is performed following computing, Obtain shown in maximum color passage transmissivity distribution map t_mc, result such as Fig. 3 (d).

$t\_mc=\frac{mc-170}{255-170}---\left(6\right)$

Step 6, transmissivity merges, it is thus achieved that the optimal transmission rate distribution to sky immunity.Color Channel transmissivity minimum to step 4 Distribution map t_dc, step 5 maximum color passage transmissivity distribution map t_mc perform point-by-point comparison and take the mixing operation of maximum, obtain Must be to shown in the optimal transmission rate distribution map t*, result such as Fig. 3 (e) of sky immunity.

Step 7: perform defogging, obtains mist elimination reconstruct image J (x), shown in final mist elimination result such as Fig. 3 (f).

$J\left(x\right)=\frac{I\left(x\right)-A}{t\left(x\right)}+A---\left(7\right).$

Claims (5)

1. a defogging method based on the variable Morphological Reconstruction of shape, it is characterised in that comprise the following steps:

Step 1: have mist image to original, asks the three-channel minimum pixel value of each point R, G, B as minimum color channel image Dc, max pixel value, as maximum color channel image mc, carries out local minimum value filtering to minimum color channel image, it is thus achieved that Rough estimate dark primary channel image DarkImg；

Step 2: determine atmosphere light position candidate from rough estimate dark primary channel image DarkImg, original have in mist image seek Correspondence is looked for have the value of maximum brightness point as air light value A；

Step 3: have the gray-scale map of mist image for guiding figure with original, calculates deformable structure element, with the structural element pair generated Minimum color channel image dc carries out adaptive median filter operation, it is thus achieved that repressed structural images dc_med of details；

Step 4: rough estimate dark primary channel image DarkImg is reconstruct mark image in step 1, with structure chart in step 3 As dc_med is reconstruct template image, carries out gray scale morphology and expand reconstruct, it is thus achieved that dark channel image DarkImg* become more meticulous； In conjunction with air light value A, utilizeCalculate the minimum Color Channel transmissivity distribution map t_dc become more meticulous；

Step 5: maximum color channel image mc is performed computingObtain maximum color passage transmissivity distribution map t_mc； T is threshold value；

Step 6: to the minimum Color Channel transmissivity distribution map t_dc become more meticulous, maximum color passage transmissivity distribution map t_mc Perform point-by-point comparison and take the mixing operation of maximum, it is thus achieved that the optimal transmission rate distribution map t* to sky immunity；

Step 7: directly utilizing air light value A and optimal transmission rate t* has mist image to carry out mist elimination sharpening to original, obtains mist elimination Reconstruct image J (x):

$J\left(x\right)=\frac{I\left(x\right)-A}{t\left(x\right)}+A---\left(7\right).$

A kind of defogging method based on the variable Morphological Reconstruction of shape the most according to claim 1, it is characterised in that step Carrying out in 1 arranging window size during local minimum value filtering is 7 pixel units.

A kind of defogging method based on the variable Morphological Reconstruction of shape the most according to claim 1, it is characterised in that step 2 particularly as follows: the pixel taking 0.1% brightness from rough estimate dark primary channel image DarkImg maximum determines atmosphere light position candidate, Searching correspondence in mist image is had to have the value of maximum brightness point as air light value A original.

A kind of defogging method based on the variable Morphological Reconstruction of shape the most according to claim 1, it is characterised in that step In 3:

d_λRepresenting deformable structure unit length, definition mode is as follows:

$\left\{\begin{array}{c}{d}_{\mathrm{\λ}}(x,x)=0\\ d_{\mathrm{\λ}}(x,y)=\underset{\mathrm{\σ}}{min}\left\{L\left(\mathrm{\σ}\right)\right\}\end{array}\right.---\left(1\right)$

Wherein, the passage length between L (σ) representative point x and some y:

$L\left(\mathrm{\σ}\right)=\underset{i=0}{\overset{n}{\mathrm{\Σ}}}\[1+\mathrm{\λ}\·d_{pixel}\left(\mathrm{Im}age\right(x_i),\mathrm{Im}age(x_{i+1}\left)\right)\]---\left(2\right)$

d_pixelFor gray scale difference value, σ=(x=x0, x1 ... xn=y) represents the passage between some x and some y, and λ is arithmetic number parameter； When λ=0, deformable structure element is the whole filter kernel region of definition, is equivalent to fixed structure element；

Select original have the gray-scale map of mist image as guiding figure；When calculating deformable structure unit, set to deformable structure unit length Put the upper limit；

A_{λ, r}=y | d_λ(x,y)≤r} (3)

R is structural elements length limit；λ=0.5；

$r=\frac{\underset{y\∈\mathrm{\Ω}\left(x\right)}{\Σ}{d}_{\mathrm{\λ},y}}{N}---\left(4\right)$

Deformable structure unit determines, minimum color channel image dc is carried out medium filtering；Obtain the repressed knot of fine texture texture Composition is as dc_med.

A kind of defogging method based on the variable Morphological Reconstruction of shape the most according to claim 1, it is characterised in that step In 5, threshold value T is 170.