CN106251296A - A kind of image defogging method and system - Google Patents

Abstract

This application discloses a kind of image defogging method and system, the method comprise the steps that and treat artwork I of mist elimination and guiding figure p carries out down-sampled；Artwork I and guiding figure p mean μ in kernel function window is asked for for down-sampled figure_k、And varianceThe Steerable filter figure q of down-sampled figure is tried to achieve based on described average and variance_iThe absorbance of the most down-sampled figure；The absorbance of described down-sampled figure is carried out liter thin absorbance t for sampling acquisition artwork；Treat mist elimination image by described thin absorbance to carry out recovering without mist image.The present invention both ensure that image detail, improves again efficiency, thus ensure that mist elimination technology practicality the most in the terminal.

Description

A kind of image defogging method and system

Technical field

The application relates to general image real time transfer field, particularly relates to a kind of image defogging method and system.

Background technology

The vile weathers such as air quality was degenerated serious in recent years, haze occur that frequently, PM2.5 value increasingly causes people's Extensive concern.The image of shooting under having greasy weather gas, owing to medium muddy in air is serious on the absorption of light and scattering impact, Make " transmission light " strength retrogression, so that the light intensity that optical pickocff receives there occurs change, directly result in image comparison Degree reduces, and reduced dynamic range is smudgy, and definition is inadequate, and image detail information is inconspicuous, and many features are capped or mould Stick with paste, can identification being substantially reduced of information.Meanwhile, color fidelity declines, and serious color displacement and distortion occurs, not only makes Visual effect is deteriorated, and image sight reduces, and also can affect the process in image later stage, gives and judges that target brings certain being stranded Difficulty, directly limits and have impact on outdoor target recognition and tracking, intelligent navigation, highway visual surveillance, Satellite Remote Sensing, military affairs The performance of the system utility such as aerial reconnaissance, causes strong influence to the everyways such as production and life.With highway monitoring it is Example, owing to having a dense fog, the visibility of road is substantially reduced, and the traffic information that driver is obtained by vision is the most inaccurate, enters The one step impact interpretation to environment, it is easy to vehicle accident occurs, now closes at a high speed or highway is restricted driving, to the trip of people Bring great inconvenience.And, the degraded image that the greasy weather obtains also result in the biggest difficulty to monitoring traffic conditions.Again than As in terms of military surveillance and remote sensing navigation, the identification of information and place are comprehended and are caused a deviation by greasy weather degraded image, it is impossible to correct Differentiate target, in some cases, limited by condition, scout and monitor that work can not repeat, so according to limited It is extremely important that image document obtains target information more accurately.Especially give the impact that airborne Visible imaging system brings Bigger, the stability of its work of severe jamming and reliability, greatly limit its application in terms of Geological Hazards Monitoring, and Geological disaster is many with vile weathers such as dense fogs before and after occurring, and produces serious influence search site rescue work.

Accordingly, it would be desirable to image mist elimination technology strengthens or repairs, to improve visual effect and to facilitate the later stage to process above-mentioned regarding Feel that the image portion information of poor effect lacks.Image mist elimination is fuzzy to have mist image by certain technology to degrade Interference with means remove mist, recovers effective image detail information, obtains the high-quality definition of satisfied visual effect The technology of image carries out a kind of technology of sharpening process.

So far, research worker mainly studies this problem of image mist elimination from both direction: based on image enhaucament Direction and direction based on physical model.The defogging method in image enhaucament direction is it is not intended that there is mist air to make image degradation degrade Actual physics process, but for the image after degrading itself, improve the contrast of image, prominent figure by image procossing way The feature of picture, improves the visual effect of image and is easy to computer vision system to the analysis of image and process.Based on physics mould The method of type shows keen interest, this direction it is crucial that physically study fog degraded image mechanism, build Go out corresponding physical model, understand fog on visual system affect essence after, then the impact of image is offset by fog and Remove.

In recent years, single image defogging method receives the concern of a lot of scholar, and these methods are by using single image In the prior information that comprises or propose some reasonably it is assumed that realize image mist elimination.Tan [1] is by maximizing local contrast Method realize mist elimination, the scene graph color that the method recovers easily tends to supersaturation, and scene restore be not built upon specific Physical model on.Fattal [2] is by assuming that body surface reflectance (Surface shading) and transmission value (Transmission) partial statistics is uncorrelated, utilize independent component analysis (Independent component analysis, ICA) scene reflectivity rate is estimated.Owing to the method is statistical properties based on input data, this statistical property is for thick fog Will lose efficacy with low signal-to-noise ratio situation.He etc. [3] propose dark channel prior defogging method first and achieve good mist elimination Effect, owing to soft stingy figure (the Soft image matting) algorithm used when transmission figure is optimized by the method can consume very Big internal memory causes processing speed slow, it is impossible to realize the real-time process of image.

All fronts all over products is upgraded to CC 2015 version in this month by Adobe company, and new edition Photoshop is mainly to being System computing is optimized, but brand-new Camera Raw 9.1 adds a very useful defogging mirror function.Pass through Actual contrast, defogging mirror is tackling in the gloomy problem of picture caused due to light problem, effect or it will be apparent that particularly For promoting in richness and the photo stereovision of color.After using defogging mirror, photo can seem sharper keen, and level is apparent, Color is more gorgeous.Finding in practice that defogging mirror is little in the effect of haze sky, or accurately say, defogging mirror can allow photo Color is more gorgeous, but for inherently not having the haze weather of color, the function of defogging mirror is also limited.Mist elimination Mirror can intensify, promote color, strengthen level, improve acutance, and this function the most simply promotes contrast or sharp Change, coordinate other regulations of later stage software, the problem that can quickly improve picture burnt hair, but can not inherently remove haze.

Summary of the invention

An object of the application is to provide a kind of and had not only ensured image detail but also improved efficiency thus ensure the image of practicality Defogging method and/or system.

An object of the application is realized by a kind of image defogging method, and the method includes:

Artwork I and the guiding figure p that treat mist elimination carry out down-sampled；

Artwork I and guiding figure p mean μ in kernel function window is asked for for down-sampled figure_k、And variance

The Steerable filter figure q of down-sampled figure is tried to achieve based on described average and variance_i, wherein q_i=a_kI_i+b_k；I is pixel tag, w_kFor kth kernel function window, | w | is window Number of pixels in Kou, ε is smoothing factor；Wherein said Steerable filter figure is corresponding to the absorbance of down-sampled figure；

The absorbance of described down-sampled figure is carried out liter thin absorbance t for sampling acquisition artwork；

Treat mist elimination image by described thin absorbance to carry out recovering, wherein without mist imageWherein J is for wanting The image without mist recovered, A is global atmosphere strength factor.

An object of the application is also realized by a kind of image mist elimination system, and this system includes:

Down-sampled module, carries out down-sampled for artwork I and guiding figure p treating mist elimination；

Average and variance obtain module, for asking for artwork I and guiding figure p in kernel function window for down-sampled figure Mean μ_k、And variance

Down-sampled figure absorbance obtains module, for trying to achieve the Steerable filter figure of down-sampled figure based on described average and variance q_i, wherein q_i=a_kI_i+b_k；I is pixel tag, w_kFor kth Kernel function window, | w | is the number of pixels in window, and ε is smoothing factor；Wherein said Steerable filter figure corresponds to down-sampled figure Absorbance；

Thin absorbance obtains module, obtains the thin transmission of artwork for the absorbance of described down-sampled figure carries out liter sampling Rate t；

First image-restoration module, is used for using described thin absorbance to treat mist elimination image and carries out recovering without mist image, its InWherein J is image without mist to be recovered, and A is global atmosphere strength factor.

In this manual, down-sampled i.e. refer to that sampling number reduces.For the image of a width N*M, if down-sampled Coefficient is k, is the most i.e. that each row and column take a some composition piece image every k point in artwork.Made by down-sampled process The size obtaining image diminishes, and preserves the information of image as far as possible.

Rise sampling to be also referred to as or image interpolation, be i.e. two-dimensional interpolation for image.If liter downsampling factor is k, i.e. exist Inserting k-1 point between artwork n and n+1 2, main purpose is enlarged drawing picture.

Present invention improvement based on dark channel prior mist elimination algorithm.For soft stingy nomography, there is higher computational load, The requirement of real-time cannot be met, utilize Steerable filter algorithm that absorbance is optimized, and use adding of Steerable filter to calculate quickly Method is accelerated so that it is meet the requirement of real-time.Specifically, first guiding figure P and artwork I are carried out down-sampled, for down-sampled Figure has sought related parameter, is reverted in artwork by a liter sampling the most again.It not that artwork is carried out the when of i.e. asking for thin absorbance Ask for, but first artwork carried out down-sampled, be such as reduced into the 1/4 of artwork, calculate the absorbance of down-sampled figure (little figure), By the way of interpolation, obtain the thin absorbance that artwork is general the most again, thus substantially increase while ensureing filter effect Performing speed, this is applied to mobile terminal device for mist elimination algorithm is significantly, because the calculating energy of mobile terminal Power is limited.Such as, by the present invention, the image of 1024*578 is carried out mist elimination process, speed can be promoted 3 times.

Accompanying drawing explanation

The present invention will be below with reference to accompanying drawing and combine preferred embodiment and illustrate more completely.

Fig. 1 is the flow chart of the embodiment according to the inventive method.

Fig. 2 is the structural representation of the embodiment according to present system.

For clarity, the figure that these accompanying drawings are schematically and simplify, they only give for understanding institute of the present invention Necessary details, and omit other details.

Detailed description of the invention

By detailed description given below, the scope of application of the present invention will be apparent to.It will be appreciated, however, that in detail While thin description and object lesson show the preferred embodiment of the present invention, they are given only for illustration purpose.

In computer vision and image processing techniques, based on atmospherical scattering model, typically with following equations, band mist is described The model of image:

I (x)=J (x) t (x)+A [1-t (x)] (1)

Wherein, I (x) treats mist elimination image exactly, and J (x) is intended to the picture rich in detail without mist recovered, and parameter A is that global atmosphere is strong Degree coefficient, t (x) is absorbance, and image mist elimination is exactly known conditions I (x), seeks target J (x).

Dark primary priori: think the outdoor natural scene image without mist, after dark primary priority handle, major part The brightness of pixel is close to zero, if there is the pixel that a large amount of brightness is higher in dark primary image, then these brightness should be from sky Fog in gas or sky, for the original image of atomization, initial saturating by obtaining from the result of dark primary priority treatment View and air intensity level.

So, for outdoor image J, dark primary is defined as:

$J^{dark}\left(x\right)=\underset{y\∈\mathrm{\Ω}\left(x\right)}{\min }\left(\underset{c\∈\{r,g,b\}}{min}{J}^c\right(y\left)\right)---\left(2\right)$

Wherein c represents each passage of coloured image, and Ω (x) represents the set of all windows centered by pixel x, y It it is a window.

Obtain according to the theory that dark primary is preferential:

J^dark→0 (3)

There is this priori, then can be carried out mathematical derivation to solve problem:

Formula (1) is carried out deformation obtain:

$\frac{I^c\left(x\right)}{A^c}=t\left(x\right)\frac{J^c\left(x\right)}{A^c}+1-t\left(x\right)---\left(4\right)$

Assume that at each window internal transmission factor t (x) be constant, be defined asAnd A value gives, to formula (4) both sides Do twice minimum operation simultaneously, then obtain:

$\underset{y\∈\mathrm{\Ω}\left(x\right)}{min}\left(\underset{c}{min}\frac{I^c\left(x\right)}{A^c}\right)=\tilde{t}\left(x\right)\underset{y\∈\mathrm{\Ω}\left(x\right)}{min}\left(\underset{c}{min}\frac{J^c\left(x\right)}{A^c}\right)+1-\tilde{t}\left(x\right)---\left(5\right)$

Derive in conjunction with formula (2), (3)

$\underset{y\∈\mathrm{\Ω}\left(x\right)}{min}\left(\underset{c\∈\{r,g,b\}}{min}\frac{J^c\left(y\right)}{A^c}\right)=0---\left(6\right)$

Bring formula (6) into formula (5), obtain treating the thick absorbance of mist elimination figure:

$\widetilde{t\left(x\right)=1-}\underset{y\∈\mathrm{\Ω}\left(x\right)}{min}\left(\underset{c\∈\{r,g,b\}}{min}\frac{I^c\left(y\right)}{A^c}\right)---\left(7\right)$

Above-mentioned inference all assume that during global atmosphere intensity A value known.In practice, general by dark figure from Have in mist image and obtain this value A.For a secondary triple channel figure, dark figure only has a width, is gray-scale map, corresponding three passages A value takes identical value.The method asking for A value is as follows:

1) from dark figure, the pixel of front 0.05%-1% such as 0.1% is taken according to the size of brightness.

2) in these positions, in the value of the original point with maximum brightness having and finding correspondence in mist image I, as A Value.

In order to obtain the finest absorbance, it is ensured that image detail, simultaneously the most again can the efficiency of boosting algorithm, it is ensured that in fact By property, the present invention proposes the method for quick Steerable filter.

Why thin absorbance is asked in selective guide filtering, on the one hand, Steerable filter has good edge and keeps flat It is sliding that (why the scenery in image can be clearly recognizable, is because between object there is the border that gray scale significantly changes.And When borderline pixel is carried out smothing filtering, the intermediate value of easy choice neighborhood or average, all can reduce limit to a certain extent The gray scale significance on boundary, thus cause the fuzzy of image.Therefore, it is intended that when image is smoothed, only to noise portion Dividing and be smoothed, maintain again the original gamma characteristic of image boundary simultaneously, this kind of wave filter is referred to as edge and keeps flat Filter slide.Several frequently seen edge preserving smooth filter device is presented herein below) characteristic, on the other hand, Steerable filter can be used to surpass The most smooth, it can make filtering output more orderly and fewer than input smooth.

Steerable filter principle

Image orientation filtering is the filtering that a linear shifting is variable, comprises artwork I, guides figure p and output image q. When initial, guide figure p equal to artwork I.So for the ith pixel in output image, its computational methods can be expressed as:

$q_i=\underset{j}{\Σ}{W}_{ij}{\left(I\right)}_{p_j}---\left(8\right)$

In formula, i and j is pixel tag, W_ijFor filtering, it is defined as:

$W_{ij}\left(I\right)=\frac{1}{|w{|}^2}\underset{k:(i,j)\∈w_k}{\Σ}(1+\frac{(I_i-{\mathrm{\μ}}_k)(I_j-{\mathrm{\μ}}_k)}{{\mathrm{\σ}}_k^2+\mathrm{\ϵ}})---\left(9\right)$

Wherein w_kFor kth kernel function window, | w | is the number of pixels in window, μ_kWithFor guiding figure in window Average and variance.ε is smoothing factor.

Owing to the model of Steerable filter meets the model of local linear, it is possible to assume input and the output of this kernel function Linear relationship is met in a two-dimentional window, it may be assumed that

q_i=a_kI_i+b_k (10)

Here a and b is the coefficient of this linear function when window center is positioned at k, and above formula both sides are taken gradient, can obtain Arrive:

Q=a I (11)

Followed by obtaining coefficient a, b of linear function, being realized by the method for linear regression, namely we pass through Method of least square makes following formula minimum:

Obtain:

$a_k=\frac{\frac{1}{\left|w\right|}{\mathrm{\Σ}}_{i\∈w_k}{I}_i{p}_i-{\mathrm{\μ}}_k{\stackrel{\‾}{p}}_k}{Q_k^2+\mathrm{\ϵ}}$

$b_k={\stackrel{\‾}{p}}_k-a_k{\mathrm{\μ}}_k$

In order to obtain a, b, need to calculate mean μ_k,Variance

Fig. 1 shows the flow chart of image defogging method according to an embodiment of the invention, and the method comprising the steps of: at S10, Artwork I and the guiding figure p that treat mist elimination carry out down-sampled, as being reduced into the 1/4 of artwork.In step S20, ask for former for down-sampled figure Figure I and guiding figure p mean μ in kernel function window_k、And varianceIn step S30, ask based on described average and variance Obtain the Steerable filter figure q of down-sampled figure_i, wherein q_i=a_kI_i+b_k； I is pixel tag, w_kFor kth kernel function window, | w | is the number of pixels in window, and ∈ is smoothing factor and takes between 0-1 Value；After obtaining a, b, just can obtain Steerable filter figure q according to formula (10).For triple channel image, need to obtain respectively The Steerable filter figure of tri-passages of R, G, B, the absorbance that its corresponding down-sampled figure is corresponding on three passages.In step S40, right The absorbance of described down-sampled figure carries out liter sampling and obtains thin absorbance t of artwork.Down-sampled figure absorbance asks for efficiency comparison Height, if the figure of 320*240, the absorbance figure size obtained is also 320*240, but if artwork is 640*480, fall is adopted The absorbance figure of master drawing directly uses and does not mate with the size of artwork, so being obtained by down-sampled figure by the method for liter sample interpolation To small size absorbance expand to the size of artwork, to mate with the size of artwork when calculating, it is simple to directly calculate.For For image, image is carried out down-sampled and up-sampling process, mean μ_k、And varianceParameter is constant.In step S50, obtains the thick absorbance of artwork based on dark channel prior mist elimination algorithm.In step S60, treat mist elimination by described thick absorbance Artwork carry out without mist image recover, whereinWherein J is the image recovering to obtain, and A is global atmosphere intensity Coefficient.Here mathematical symbol I is the abbreviation of I (x), and A is A (x), t and is t (x), J and is J (x).In step S70, right Carry out the image obtained without the recovery of mist image by thick absorbance, carry out again once recovering without mist image by described thin absorbance, To picture rich in detail without mist to be recovered.The method of this embodiment reduces amount of calculation, improves efficiency, ensures not affect figure simultaneously The details of sheet.

Fig. 2 shows the schematic diagram of image mist elimination system according to an embodiment of the invention, and this system includes: down-sampled mould Block 10, carries out down-sampled for artwork I and guiding figure p treating mist elimination；Average and variance obtain module 20, for adopting for fall Master drawing asks for artwork I and guiding figure p mean μ in kernel function window_k、 And varianceDown-sampled figure absorbance obtains Module 30, for trying to achieve the Steerable filter figure q of down-sampled figure based on described average and variance_i, wherein q_i=a_kI_i+b_k； I is pixel tag, w_kFor kth kernel function window, | w | is window Number of pixels in Kou, ∈ is smoothing factor；Wherein said Steerable filter figure is corresponding to the absorbance of down-sampled figure；Thin transmission Rate obtains module 40, obtains thin absorbance t of artwork for the absorbance of described down-sampled figure carries out liter sampling；Air intensity Coefficient obtains module 50, is used for obtaining described global atmosphere strength factor A；Thick absorbance obtains module 60, for based on helping secretly Road priori mist elimination algorithm obtains the thick absorbance of artwork；Second image-restoration module 70, is used for using described thick absorbance to treat The artwork of mist elimination carries out recovering without mist image；First image-restoration module 80, is used for using described thin absorbance to the second image The image that recovery module obtains carries out recovering, wherein without mist image againWherein J is to be recovered clear without mist Clear image, A is global atmosphere strength factor.In an embodiment, air strength factor module includes: pixel obtains submodule 52, The pixel of front 0.05%-1% is taken for size according to brightness from the dark figure treat mist elimination image；Maximum brightness point obtains Submodule 54, in the location of pixels taken out, finds the corresponding point with maximum brightness in treating mist elimination image Value is as A value.

Unless explicitly stated otherwise, singulative " ", " being somebody's turn to do " as used herein all include that plural reference (i.e. has " at least one " The meaning).It will be further understood that terminology used herein " has ", " including " and/or " comprising " shows that existence is described Feature, step, operation, element and/or parts, but do not preclude the presence or addition of other features one or more, step, operation, Element, parts and/or a combination thereof.Term "and/or" includes any of one or more relevant item enumerated as used in this And all combinations.Unless explicitly stated otherwise, the step of any method disclosed herein need not accurately perform according to disclosed order.

Some preferred embodiments are in explanation made above, it should be emphasized, however, that the present invention is not limited to this A little embodiments, but can realize with the alternate manner in the range of present subject matter.

List of references

[1]Tan R T.Visibility in bad weather from a single image.In: Proceedings of the 2008 IEEE Conference on Computer Vision and Pattern Recognition.Anchorage,AK,USA:IEEE,2008.1-8

[2]Fattal R.Single image dehazing.ACM Transactions on Graphics,2008, 27(3):Article No.72

[3]He K M,Sun J,Tang X O.Single image haze removal using dark channel prior.In:Proceedings of the 2009 IEEE Conference on Computer Vision and Pattern Recognition.Miami,USA:IEEE,2009.1956-1963

Claims (10)

1. an image defogging method, it is characterised in that described method includes:

Artwork I and the guiding figure p that treat mist elimination carry out down-sampled；

Artwork I and guiding figure p mean μ in kernel function window is asked for for down-sampled figure_k、And variance

The Steerable filter figure q of down-sampled figure is tried to achieve based on described average and variance_i, wherein q_i=a_kI_i+b_k；I is pixel tag, w_kFor kth kernel function window, | w | is window Number of pixels in Kou, ε is smoothing factor；Wherein said Steerable filter figure is corresponding to the absorbance of down-sampled figure；

The absorbance of described down-sampled figure is carried out liter thin absorbance t for sampling acquisition artwork；

Treat mist elimination image by described thin absorbance to carry out recovering, wherein without mist imageWherein J is to be recovered Image without mist, A is global atmosphere strength factor.

Method the most according to claim 1, it is characterised in that described method also includes:

The thick absorbance of artwork is obtained based on dark channel prior mist elimination algorithm；

The artwork treating mist elimination by described thick absorbance carries out recovering without mist image；

Wherein said with described thin absorbance treat mist elimination image carry out without mist image recover include with described thin absorbance to lead to Cross and carry out mist elimination by thick absorbance and process the image that obtains and again carry out without the recovery of mist image.

Method the most according to claim 1, it is characterised in that described global atmosphere strength factor A passes through below step Try to achieve:

Take the pixel of front 0.05%-1% according to the size of brightness from the dark figure treating mist elimination image；

In the location of pixels taken out, in treating mist elimination image, find the value of the corresponding point with maximum brightness as A value.

Method the most according to claim 3, it is characterised in that big according to brightness from the dark figure treating mist elimination image Little take front 0.1% pixel.

Method the most according to claim 1, it is characterised in that described smoothing factor takes the value between 0～1.

6. an image mist elimination system, it is characterised in that described system includes:

Down-sampled module, carries out down-sampled for artwork I and guiding figure p treating mist elimination；

Average and variance obtain module, for asking for artwork I and guiding figure p average in kernel function window for down-sampled figure μ_k、And variance

Down-sampled figure absorbance obtains module, for trying to achieve the Steerable filter figure q of down-sampled figure based on described average and variance_i, its Middle q_i=a_kI_i+b_k；I is pixel tag, w_kFor kth core letter Number window, | w | is the number of pixels in window, and ε is smoothing factor；Wherein said Steerable filter figure is saturating corresponding to down-sampled figure Penetrate rate；

Thin absorbance obtains module, obtains thin absorbance t of artwork for the absorbance of described down-sampled figure carries out liter sampling；

First image-restoration module, is used for using described thin absorbance to treat mist elimination image and carries out recovering, wherein without mist imageWherein J is image without mist to be recovered, and A is global atmosphere strength factor.

System the most according to claim 6, it is characterised in that described system also includes:

Thick absorbance obtains module, for obtaining the thick absorbance of artwork based on dark channel prior mist elimination algorithm；

Second image-restoration module, the artwork for using described thick absorbance to treat mist elimination carries out recovering without mist image；

Wherein said first image-restoration module includes the figure obtained described second image-restoration module by described thin absorbance As again carrying out recovering without mist image.

System the most according to claim 6, it is characterised in that described system also includes that air strength factor obtains module, For obtaining described global atmosphere strength factor A；Wherein said air strength factor module includes:

Pixel obtains submodule, for taking front 0.05%-1%'s according to the size of brightness from the dark figure treating mist elimination image Pixel；

Maximum brightness point obtains submodule, in the location of pixels taken out, finds the tool of correspondence in treating mist elimination image There is the value of point of maximum brightness as A value.

System the most according to claim 8, it is characterised in that described pixel obtains submodule from treating helping secretly of mist elimination image Road figure takes the pixel of front 0.1% according to the size of brightness.

System the most according to claim 6, it is characterised in that described smoothing factor takes the value between 0～1.