CN103500440A - Method for eliminating cloud and haze of atmospheric degraded image - Google Patents

Abstract

The invention discloses a method for eliminating cloud and haze of an atmospheric degraded image. The method comprises the steps of the acquisition of a dark channel image and mid-value filtering, the self-adaptive degradation and acquisition of an atmospheric light image, the detailing of an atmospheric transmission function and the visual correction of a color domain. The atmospheric degraded image to be processed by the method has the problems of different illumination intensities, contrast ratios and dynamic ranges, a dark channel theory model is adopted, and the mid-value filtering of the dark channel image, the self-adaptive degradation and acquisition of the atmospheric light image and the visual correction of the color domain are also combined to eliminate the cloud and haze, and the disadvantages in the prior art are overcome. Besides, the method for eliminating the cloud and haze of the atmospheric degraded image is simple to operate and has good application prospect.

Description

A kind of method of removing the Atmospheric Degraded Image cloud and mist

Technical field

The invention belongs to computing machine and technical field of image processing, relate in particular to a kind of method of removing the Atmospheric Degraded Image cloud and mist.

Background technology

It is the basis of the middle and high stage of Computer Image Processing improving picture quality that Atmospheric Degraded Image under visual light imaging goes the cloud and mist technology, it is intended to recover the raw information of image from the atmospheric degradation phenomenons such as cloud and mist and utmost point low-light (level), and the efficient recovery that has of image information is being brought into play significant role to the lifting of the road safety monitoring under IFR conditions, vehicle security drive, optical weapon fighting efficiency, effective supervision in enemy and we battlefield ^[1]yet, the Visible imaging system of existing Physical modeling based is subject to having a strong impact on of atmospheric conditions, under the atmospheric degradation phenomenons such as cloud and mist, the a large amount of small water droplet suspended in atmosphere, aerocolloidal decay and scattering process make atmospheric visibility, brightness degradation, the picture quality that causes the visual light imaging sensor to collect reduces, and greatly affects and limited the function of outdoor Visible imaging system under the complicated meteorology environment ^{[2] [3]}, caused afterwards domestic and international many researchers' great interest, become in recent years the forward position focus that the pattern process computer technical field receives much concern.

According to algorithm characteristic, image for the atmospheric degradation phenomenon goes the cloud and mist technology mainly to form both direction: the method based on the figure image intensifying and the method based on image restoration, method based on the figure image intensifying can only be relative the quality of lifting atmospheric degradation phenomenon hypograph, can not realize the cloud and mist that goes truly, method based on image restoration depends on the foundation of image degradation model, by atmospheric attenuation and ambient lighting are carried out to modeling, and according to strong hypothesis and prior imformation, the non-well-posedness of model solution is converted to well-posedness, thereby realize parameter analysis and the acquisition of removing the cloud and mist image, it is forward position and the focus of research both at home and abroad so far that this research direction starts from CVPR2008, especially dark primary (the Dark Channel Prior with what bright grade of happy of CVPR2009 proposed, DCP) be representative, the method possesses physics validity, but when the scene target when very large zone and atmosphere light are very similar in essence, and while not having shade to project on object, its physical model will be invalid, the present invention is from the image restoration angle, be intended to recover original high quality from the model that degrades, when the process that degrades is carried out to the adaptability physical modeling, introduce the color space correction mechanism, make the color performance abundanter, thereby improve the visual effect of image, in order to make it to have applicability widely, to make brightness of image when processing, contrast and color interrelate by the inherent bottom factor of model and image respectively, setup parameter that so just need not be artificial, and make to restore intrinsic propesties and the human eye vision perception that result meets image more, in a word, based on image restoration, go the cloud and mist method a lot, key is an optimum solution that searches out its ill degeneration physical model, thereby realize agreeing with of Neighbourhood feature and global characteristics.

Theoretical (the Dark Channel Prior of the dark primary that He Kaiming etc. propose, DCP) think, clearly on image except the sky dummy section, in the RGB Color Channel, have at least a passage to have very low intensity level, on the cloud and mist image, the intensity level of dark primary mainly is comprised of atmosphere light, the method is directly estimated transmission diagram with dark primary, and use the method for image mending to carry out smooth operation to transmission diagram, use the transmission diagram after repairing can recover image clearly, and therefrom obtain the depth map of Misty Image ^[5], concrete implementation is as follows:

At first suppose that atmosphere light A is given, further supposition is invariable at the propagation in atmosphere function of a regional area, and the atmospheric scattering model use of McCarney is got to minimum operational symbol, and same divided by A, obtains:

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

Use minimum computing in three Color Channels, have:

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

According to the rule of dark primary priori, without the dark primary item J of mist natural image ^darkshould be close to 0:

$J^{\mathrm{dark}}\left(x\right)=\underset{C}{\min }\left(\underset{y\∈\mathrm{\Ω}}{\min }\left(J^C\left(y\right)\right)\right)=0$

Due to A ^ctotal is positive number, derives:

$\underset{C}{\min }\left(\underset{y\∈\mathrm{\Ω}}{\min }\left(\frac{J^C\left(y\right)}{A^C}\right)\right)=0$

Therefore, can estimate simply transmissivity t:

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

If remove up hill and dale the existence of mist, it is untrue that image can seem, and depth perception can lose, so can, by introduce a constant ω (0<ω≤1) in above formula, retain the mist that a part covers remote scenery:

$\tilde{t}\left(x\right)=1-\underset{C}{w\min }\left(\underset{y\&Element;\mathrm{\&Omega;}}{\min }\left(\frac{I^C\left(y\right)}{A^C}\right)\right)$

It is rough estimating transmissivity by above formula, and in order to improve precision, application softmatting algorithm improves the transmissivity distribution function, and the transmittance function of note after improving is t (x), by solving following formula, obtains:

$(L+\mathrm{\&lambda;U})t=\mathrm{\&lambda;}\tilde{t}$

λ is a corrected parameter, and L is Laplce's correction matrix that Levin proposes, and by following formula, calculates the image J (x) after mist elimination:

$J\left(x\right)=\frac{I\left(x\right)-A}{\max (t\left(x\right),t_0)}+A$

The method of estimation of atmosphere light A is: first get dark primary J ^darkin the pixel of 0.1% brightness maximum, then get the value of the maximal value of these pixel correspondences in former figure as A, respectively the two width images that have slight and dense cloud and mist phenomenon are processed.

This scheme is based upon on the dark primary a priori assumption, but when very similar or cloud and mist is deep in essence with atmosphere light in very large zone when target in reality scene, just is difficult to meet this hypothesis, can make obtain help that image border is coarse, the level fuzzy secretly; The cloud and mist quality is removed in the fast size impact of son; Atmosphere illumination can not reflect true scene, and does not possess local characteristics; Cause the propagation in atmosphere function localized distortion of estimation serious; Color distortion appears in restoration result, has affected the visual effect of image; The Integral Restoration result is darker, and color is distinct not.

Summary of the invention

The purpose of the embodiment of the present invention is to provide a kind of method of removing the Atmospheric Degraded Image cloud and mist, be intended to solve that the prior art scheme exists when very similar or cloud and mist is deep in essence with atmosphere light in very large zone when target, just be difficult to meet this hypothesis, can make obtain help that image border is coarse, the level fuzzy secretly; The cloud and mist quality is removed in the fast size impact of son; Atmosphere illumination can not reflect true scene, and does not possess local characteristics; Cause the propagation in atmosphere function localized distortion of estimation serious; Color distortion appears in restoration result, has affected the visual effect of image; The Integral Restoration result is darker, the problem that color is distinct not.

The embodiment of the present invention is achieved in that a kind of method of removing the Atmospheric Degraded Image cloud and mist, and the method for described removal Atmospheric Degraded Image cloud and mist comprises the following steps:

Help secretly image obtain and the adaptive decomposition of medium filtering, atmosphere light image obtains;

The refinement of propagation in atmosphere function;

The vision correction of color gamut.

Further, help image secretly and generate to adopt the image adaptive piecemeal to process thought, obtain the corresponding image I of helping secretly of original image helping secretly on priori theoretical foundation _d(x, y), and use median filter smoothness of image to process, available following formula defines the image I of helping secretly generated _d(x, y):

$I_D(x,y)=\underset{c\&Element;\{r,g,b\}}{\min }\left(\underset{\mathrm{\&Omega;}}{\min }\left(I_c(x,y)\right)\right)$

Wherein, I _cthe dedicated tunnel that (x, y) is original image; Ω is adaptive image subblock, is a square area centered by (x, y), for the coarse phenomenon that prevents that depth jump from producing, avoids occurring gridiron pattern effect and halation vestige in result, need carry out medium filtering to helping image secretly;

Further, the atmosphere light image is obtained and is adopted the self-adaption two-dimensional empirical mode decomposition, after carrying out 5 decomposition, obtains original image corresponding atmosphere light image A (x, y), and carrys out even local atmosphere light by low-pass filtering.

Further, the self-adaption two-dimensional empirical mode decomposition, concrete implementation step is as follows:

Step 1, by the local extremum point in the geodetic operator identification image in mathematical morphology;

Step 2, the Interpolation Property of Radial Basis Function method structure envelope surface that employing has been optimized, maximum point and minimum point are carried out respectively to interpolation arithmetic, obtain maximum point envelope surface and minimum point envelope surface after computing, two curved surface datas are averaging and obtain average envelope surface data, therefore, the upper lower envelope Solve problems in image space just changes into the discrete data point interpolation reconstruction problem of three-dimension curved surface, and solving the envelope plane can be expressed as:

$s\left(x\right)=c_0+c_{1}x+c_{2}y+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_i\mathrm{\&phi;}\left(\right||x-x_i|\left|\right)$

Wherein, s (x) is interpolation knot, c ₀, c ₁, c ₂, λ _ifor multinomial coefficient and radial basis function combination coefficient, ‖ ‖ is euclideam norm, and φ () is radial basis function, by solving N+3 unit large linear systems once, solves c ₀, c ₁, c ₂, λ _ivalue, and then substitution coordinate figure a little obtain whole interpolation curved surface, to the methods such as thin-plate, multiquadratics of having chosen of φ (), improve being chosen for of φ ():

$\mathrm{\&phi;}\left(R\right)=\left\{\begin{array}{cc}{(1-R)}^P\mathrm{PR},& R<1\\ 0,& R>1\end{array}\right.$

Wherein, R=‖ x-x _i‖ is Euclidean distance, and P is a constant coefficient;

Step 3, deduct the average envelope surface with former curved surface;

Step 4, judge whether to meet end condition, because accumulateing the number of modular function zero crossing in two dimension can't add up, so can be by the constraint condition of accumulateing modular function in two dimension during Bidimensional Empirical Mode Decomposition the stop condition as screening process, also can be by the efficient Cauchy-type condition of convergence:

$\mathrm{SD}=\underset{x}{\overset{M}{\mathrm{\&Sigma;}}}\underset{y}{\overset{N}{\mathrm{\&Sigma;}}}\frac{{|f_{k-1}(x,y)-f_k(x,y)|}^2}{f_{k-1}^2(x,y)}$

Wherein, f _kthe pixel value of (x, y) point on image when (x, y) is k layer self-adapting Bidimensional Empirical Mode Decomposition, and can to make SD be the number between 0.2 to 0.3;

Repeating step (one)～(three), until meet given end condition, obtain accumulateing modular function image bimf in the 1st layer of two dimension ₁(x, y), deduct bimf with original image ₁(x, y) obtain the 1st layer of residual image residue1, to residue1 repeating step ()～(four), obtain successively accumulateing modular function image and N layer residual image in the N layer two dimension of image, in said process, extreme point solves, the stop condition of planar interpolation and screening is the core of this algorithm, is generally to carry out 5 times to decompose, just can obtain pure atmosphere light image, so final result is expressed as:

$r_5(x,y)=f(x,y)-\underset{k=1}{\overset{5}{\mathrm{\&Sigma;}}}{\mathrm{bimf}}_k(x,y)$

Wherein, bimf _k(x, y) accumulates the modular function image in k layer two dimension, r ₅(x, y) be the trend map picture after 5 layers of decomposition, some Atmospheric Degraded Images in reality, its atmosphere illumination is difficult to meet the local flatness of image, in order to overcome this shortcoming, improve the robustness of algorithm, can adopt the low frequency illuminance information in low-pass filtering smoothing processing atmosphere light image.

Further, the refinement of propagation in atmosphere function adopts to help secretly most processes thought, by following formula:

$t_1(x,y)=1-w\underset{\mathrm{\&Omega;}}{\min }\left(\frac{I_D(x,y)}{A(x,y)}\right)$

Obtain rough propagation in atmosphere function t ₁, wherein Ω is adaptive image subblock, and w is the parameter that the validity in order to keep restoring rear image is introduced, and value between 0 to 1, get 0.95 in experiment usually;

Adopt the softmatting algorithm, to rough propagation in atmosphere function t ₁carry out the Laplacian Matrix correction, the energy equation below optimization carrys out refinement t ₁:

E(t)=t ^TLt+λ(t-t ₁) ^TU(t-t ₁)

Obtain the propagation in atmosphere function t of refinement, wherein λ is normalized parameter, and U is and the unit matrix of the equal size of image that L is Laplce's correction matrix, above-mentioned energy equation is carried out to sparse linear and mean:

(L+λU)t=λt ₁

Wherein, λ is a very little numerical value, is set as 10 in experiment ^-4; The a certain pixel of Laplce's correction matrix L (i, j) can be expressed as:

$\underset{k|(i,j)\&Element;w_k}{\mathrm{\&Sigma;}}\{{\mathrm{\&delta;}}_{\mathrm{ij}}-\frac{1}{\left|w_k\right|}[1+{(I_i-{\mathrm{\&mu;}}_k)}^T{({\mathrm{\&Sigma;}}_k+\frac{\mathrm{\&epsiv;}}{\left|w_k\right|}{U}_3)}^{-1}(I_j-{\mathrm{\&mu;}}_k)\left]\right\}$

Wherein, δ _ijit is Kronecker function; μ _kand Σ _kwindow w _kthe average of middle pixel and covariance matrix; | w _k| be window w _kthe quantity of middle pixel; ε is normalized parameter; U ₃3 to take advantage of 3 unit matrix; When carrying out computing, image array is launched by column vector, be converted to the vector of one dimension, I _iand I _jbe designated as the value of i and j pixel under meaning in one-dimensional vector;

Further, go the correction concrete operations of cloud and mist image color territory to be, to the R of image, G, B three primary colors passage is gone respectively the cloud and mist computing:

$J_c(x,y)=\frac{I_c(x,y)-A(x,y)}{\max (t(x,y),t_0)}+A(x,y),c\&Element;\{r,g,b\}$

Wherein, I _cthe dedicated tunnel that (x, y) is original image, t ₀for the atmospheric dissipation function floor value of setting, usually get 0.1, very strong correlativity is arranged between the RGB three-component, four groups of images that visual effect is good have been added up based on Imagine Macmillan storehouse, tri-component correlation matrixs of RGB have been obtained, in its Image Fusion, as desired value, the related coefficient of any two component X and Y is defined as:

$r=\frac{{\mathrm{\&Sigma;}}_m{\mathrm{\&Sigma;}}_n(X_{\mathrm{mn}}-\stackrel{\&OverBar;}{X})(Y_{\mathrm{mn}}-\stackrel{\&OverBar;}{Y})}{\sqrt{{\mathrm{\&Sigma;}}_m{\mathrm{\&Sigma;}}_n{(X_{\mathrm{mn}}-\stackrel{\&OverBar;}{X})}^2{\mathrm{\&Sigma;}}_m{\mathrm{\&Sigma;}}_n{(Y_{\mathrm{mn}}-\stackrel{\&OverBar;}{Y})}^2}}$

Wherein:

for X, the Y average, obviously-1≤r≤1, so there is a kind of thinking in color correction: seek a linear transformation, the RGB component of image is converted.

Further, to arbitrary pixel m in the image J (x, y) that removes cloud and mist _xy=[m _r, m _g, m _b] ^t, carry out following rgb color territory and proofread and correct,

$\left\{\begin{array}{c}{n}_{\mathrm{xy}}=B^T{m}_{\mathrm{xy}}\\ C_n=B^T{C}_{m}B\end{array}\right.$

Wherein, n _xyfor proofreading and correct rear pixel vector, B is the color gamut correction matrix, C _m, C _nbe respectively and proofread and correct front and back image color interchannel covariance matrix, in order to solve color correction matrix B, can be first to positive definite matrix C _m, C _ncarry out the cholesky decomposition, obtain C _m=Q _m ^tq _m, C _n=Q _n ^tq _n, easily try to achieve B=Q _m ^-1q _n, establishing the correlation matrix that between component, related coefficient forms is R _n, easily know R _nwith C _nthere is following relation:

$C_n=\left[\begin{array}{ccc}{\mathrm{\&sigma;}}_{n1}& 0& 0\\ 0& {\mathrm{\&sigma;}}_{n2}& 0\\ 0& 0& {\mathrm{\&sigma;}}_{n3}\end{array}\right]R_n{\left[\begin{array}{ccc}{\mathrm{\&sigma;}}_{n1}& 0& 0\\ 0& {\mathrm{\&sigma;}}_{n2}& 0\\ 0& 0& {\mathrm{\&sigma;}}_{n3}\end{array}\right]}^T$

C for image to be corrected _mcan pass through

try to achieve, wherein

perhaps also can try to achieve according to relation in above formula;

Further, going to cloud and mist image color territory to proofread and correct adopts average plus-minus variance as correlation matrix and R between two other component _nobtain respectively three width images after correction, be averaging afterwards and obtain last correcting image.

The method of removal Atmospheric Degraded Image cloud and mist provided by the invention, help reason opinion model secretly by employing, and incorporate the medium filtering of helping image secretly, the adaptive decomposition of atmosphere light image obtains with the vision correction of color gamut and is removed cloud and mist, after processing, medium filtering helps image boundary saltus step exquisiteness secretly, intra-zone is level and smooth, the atmosphere light image overall variation of obtaining through adaptive decomposition is level and smooth, there is its illumination feature separately part, propagation in atmosphere function details through the softmatting thinning processing is obvious, stereovision is good, the brightness of image of proofreading and correct finally by the rgb color territory is moderate, color is saturated, details is clear, abundant information is applicable to the human eye evaluation.In addition, treatment scheme of the present invention is comparatively reasonable, and the links of processing is indispensable, absolutely proves that the present invention compresses at dynamic range of images, details highlights and recover the ability on color information.

The accompanying drawing explanation

Fig. 1 is the process flow diagram of the method for the removal Atmospheric Degraded Image cloud and mist that provides of the embodiment of the present invention;

Fig. 2 is one deck decomposition process schematic diagram of the ABEMD that provides of the embodiment of the present invention.

Embodiment

In order to make purpose of the present invention, technical scheme and advantage clearer, below in conjunction with embodiment, the present invention is further elaborated.Should be appreciated that specific embodiment described herein, only in order to explain the present invention, is not intended to limit the present invention.

Fig. 1 shows the method flow of removal Atmospheric Degraded Image cloud and mist provided by the invention.For convenience of explanation, only show part related to the present invention.

The method of removal Atmospheric Degraded Image cloud and mist of the present invention, the method for this removal Atmospheric Degraded Image cloud and mist comprises the following steps:

Help secretly image obtain and the adaptive decomposition of medium filtering, atmosphere light image obtains;

The refinement of propagation in atmosphere function;

The vision correction of color gamut.

As a prioritization scheme of the embodiment of the present invention, help image secretly and generate to adopt the image adaptive piecemeal to process thought, obtain the corresponding image I of helping secretly of original image helping secretly on priori theoretical foundation _d(x, y), and use median filter smoothness of image to process, available following formula defines the image I of helping secretly generated _d(x, y):

$I_D(x,y)=\underset{c\&Element;\{r,g,b\}}{\min }\left(\underset{\mathrm{\&Omega;}}{\min }\left(I_c(x,y)\right)\right)$

Wherein, I _cthe dedicated tunnel that (x, y) is original image; Ω is adaptive image subblock, is a square area centered by (x, y), for the coarse phenomenon that prevents that depth jump from producing, avoids occurring gridiron pattern effect and halation vestige in result, need carry out medium filtering to helping image secretly;

A prioritization scheme as the embodiment of the present invention, the atmosphere light image is obtained and is adopted the self-adaption two-dimensional empirical mode decomposition, after carrying out 5 decomposition, obtains the corresponding atmosphere light image of original image A (x, y), and by low-pass filtering carry out even local atmosphere light.

As a prioritization scheme of the embodiment of the present invention, the self-adaption two-dimensional empirical mode decomposition, concrete implementation step is as follows:

Step 1, by the local extremum point in the geodetic operator identification image in mathematical morphology;

Step 2, the Interpolation Property of Radial Basis Function method structure envelope surface that employing has been optimized, maximum point and minimum point are carried out respectively to interpolation arithmetic, obtain maximum point envelope surface and minimum point envelope surface after computing, two curved surface datas are averaging and obtain average envelope surface data, therefore, the upper lower envelope Solve problems in image space just changes into the discrete data point interpolation reconstruction problem of three-dimension curved surface, and solving the envelope plane can be expressed as:

$s\left(x\right)=c_0+c_{1}x+c_{2}y+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_i\mathrm{\&phi;}\left(\right||x-x_i|\left|\right)$

Wherein, s (x) is interpolation knot, c ₀, c ₁, c ₂, λ _ifor multinomial coefficient and radial basis function combination coefficient, ‖ ‖ is euclideam norm, and φ () is radial basis function, by solving N+3 unit large linear systems once, solves c ₀, c ₁, c ₂, λ _ivalue, and then substitution coordinate figure a little obtain whole interpolation curved surface, to the methods such as thin-plate, multiquadratics of having chosen of φ (), improve being chosen for of φ ():

$\mathrm{\&phi;}\left(R\right)=\left\{\begin{array}{cc}{(1-R)}^P\mathrm{PR},& R<1\\ 0,& R>1\end{array}\right.$

Wherein, R=‖ x-x _i‖ is Euclidean distance, and P is a constant coefficient;

Step 3, deduct the average envelope surface with former curved surface;

Step 4, judge whether to meet end condition, because accumulateing the number of modular function zero crossing in two dimension can't add up, so can be by the constraint condition of accumulateing modular function in two dimension during Bidimensional Empirical Mode Decomposition the stop condition as screening process, also can be by the efficient Cauchy-type condition of convergence:

$\mathrm{SD}=\underset{x}{\overset{M}{\mathrm{\&Sigma;}}}\underset{y}{\overset{N}{\mathrm{\&Sigma;}}}\frac{{|f_{k-1}(x,y)-f_k(x,y)|}^2}{f_{k-1}^2(x,y)}$

Wherein, f _kthe pixel value of (x, y) point on image when (x, y) is k layer self-adapting Bidimensional Empirical Mode Decomposition, and can to make SD be the number between 0.2 to 0.3;

Repeating step (one)～(three), until meet given end condition, obtain accumulateing modular function image bimf in the 1st layer of two dimension ₁(x, y), deduct bimf with original image ₁(x, y) obtain the 1st layer of residual image residue1, to residue1 repeating step ()～(four), obtain successively accumulateing modular function image and N layer residual image in the N layer two dimension of image, in said process, extreme point solves, the stop condition of planar interpolation and screening is the core of this algorithm, is generally to carry out 5 times to decompose, just can obtain pure atmosphere light image, so final result is expressed as:

$r_5(x,y)=f(x,y)-\underset{k=1}{\overset{5}{\mathrm{\&Sigma;}}}{\mathrm{bimf}}_k(x,y)$

Wherein, bimf _k(x, y) accumulates the modular function image in k layer two dimension, r ₅(x, y) be the trend map picture after 5 layers of decomposition, some Atmospheric Degraded Images in reality, its atmosphere illumination is difficult to meet the local flatness of image, in order to overcome this shortcoming, improve the robustness of algorithm, can adopt the low frequency illuminance information in low-pass filtering smoothing processing atmosphere light image.

As a prioritization scheme of the embodiment of the present invention,

Obtain rough propagation in atmosphere function t ₁, wherein Ω is adaptive image subblock, and w is the parameter that the validity in order to keep restoring rear image is introduced, and value between 0 to 1, get 0.95 in experiment usually;

Adopt the softmatting algorithm, to rough propagation in atmosphere function t ₁carry out the Laplacian Matrix correction, the energy equation below optimization carrys out refinement t ₁:

E(t)=t ^TLt+λ(t-t ₁) ^TU(t-t ₁)

Obtain the propagation in atmosphere function t of refinement, wherein λ is normalized parameter, and U is and the unit matrix of the equal size of image that L is Laplce's correction matrix, above-mentioned energy equation is carried out to sparse linear and mean:

(L+λU)t=λt ₁

Wherein, λ is a very little numerical value, is set as 10 in experiment ^-4; The a certain pixel of Laplce's correction matrix L (i, j) can be expressed as:

$\underset{k|(i,j)\&Element;w_k}{\mathrm{\&Sigma;}}\{{\mathrm{\&delta;}}_{\mathrm{ij}}-\frac{1}{\left|w_k\right|}[1+{(I_i-{\mathrm{\&mu;}}_k)}^T{({\mathrm{\&Sigma;}}_k+\frac{\mathrm{\&epsiv;}}{\left|w_k\right|}{U}_3)}^{-1}(I_j-{\mathrm{\&mu;}}_k)\left]\right\}$

Wherein, δ _ijit is Kronecker function; μ _kand Σ _kwindow w _kthe average of middle pixel and covariance matrix; | w _k| be window w _kthe quantity of middle pixel; ε is normalized parameter; U ₃3 to take advantage of 3 unit matrix; When carrying out computing, image array is launched by column vector, be converted to the vector of one dimension, I _iand I _jbe designated as the value of i and j pixel under meaning in one-dimensional vector;

As a prioritization scheme of the embodiment of the present invention, go the correction concrete operations of cloud and mist image color territory to be, to the R of image, G, B three primary colors passage is gone respectively the cloud and mist computing:

$J_c(x,y)=\frac{I_c(x,y)-A(x,y)}{\max (t(x,y),t_0)}+A(x,y),c\&Element;\{r,g,b\}$

Wherein, I _cthe dedicated tunnel that (x, y) is original image, t ₀for the atmospheric dissipation function floor value of setting, usually get 0.1, very strong correlativity is arranged between the RGB three-component, four groups of images that visual effect is good have been added up based on Imagine Macmillan storehouse, tri-component correlation matrixs of RGB have been obtained, in its Image Fusion, as desired value, the related coefficient of any two component X and Y is defined as:

$r=\frac{{\mathrm{\&Sigma;}}_m{\mathrm{\&Sigma;}}_n(X_{\mathrm{mn}}-\stackrel{\&OverBar;}{X})(Y_{\mathrm{mn}}-\stackrel{\&OverBar;}{Y})}{\sqrt{{\mathrm{\&Sigma;}}_m{\mathrm{\&Sigma;}}_n{(X_{\mathrm{mn}}-\stackrel{\&OverBar;}{X})}^2{\mathrm{\&Sigma;}}_m{\mathrm{\&Sigma;}}_n{(Y_{\mathrm{mn}}-\stackrel{\&OverBar;}{Y})}^2}}$

Wherein:

for X, the Y average, obviously-1≤r≤1, so there is a kind of thinking in color correction: seek a linear transformation, the RGB component of image is converted.

As a prioritization scheme of the embodiment of the present invention, to arbitrary pixel m in the image J (x, y) that removes cloud and mist _xy=[m _r, m _g, m _b] ^t, carry out following rgb color territory and proofread and correct,

$\left\{\begin{array}{c}{n}_{\mathrm{xy}}=B^T{m}_{\mathrm{xy}}\\ C_n=B^T{C}_{m}B\end{array}\right.$

Wherein, n _xyfor proofreading and correct rear pixel vector, B is the color gamut correction matrix, C _m, C _nbe respectively and proofread and correct front and back image color interchannel covariance matrix, in order to solve color correction matrix B, can be first to positive definite matrix C _m, C _ncarry out the cholesky decomposition, obtain C _m=Q _m ^tq _m, C _n=Q _n ^tq _n, easily try to achieve B=Q _m ^-1q _n, establishing the correlation matrix that between component, related coefficient forms is R _n, easily know R _nwith C _nthere is following relation:

$C_n=\left[\begin{array}{ccc}{\mathrm{\&sigma;}}_{n1}& 0& 0\\ 0& {\mathrm{\&sigma;}}_{n2}& 0\\ 0& 0& {\mathrm{\&sigma;}}_{n3}\end{array}\right]R_n{\left[\begin{array}{ccc}{\mathrm{\&sigma;}}_{n1}& 0& 0\\ 0& {\mathrm{\&sigma;}}_{n2}& 0\\ 0& 0& {\mathrm{\&sigma;}}_{n3}\end{array}\right]}^T$

C for image to be corrected _mcan pass through

try to achieve, wherein

perhaps also can try to achieve according to relation in above formula;

As a prioritization scheme of the embodiment of the present invention, go to cloud and mist image color territory to proofread and correct and adopt average plus-minus variance as correlation matrix and R between two other component _nobtain respectively three width images after correction, be averaging afterwards and obtain last correcting image.

Below in conjunction with drawings and the specific embodiments, application principle of the present invention is further described.

As shown in Figure 1, the method for the removal Atmospheric Degraded Image cloud and mist of the embodiment of the present invention comprises the following steps:

S101: help secretly image obtain and the adaptive decomposition of medium filtering, atmosphere light image obtains;

S102: the refinement of propagation in atmosphere function;

S103: the vision correction of color gamut.

The concrete steps of the embodiment of the present invention are:

The first step, help image secretly and generate: adopt the image adaptive piecemeal to process thought, obtain the corresponding image I of helping secretly of original image helping secretly on priori theoretical foundation _d(x, y), and use median filter smoothness of image to process,

In the present invention, physical model used is the atmospheric scattering model of McCarney, and this model is widely used in the pattern process computer field, and the physical characteristics of transmitting under the atmospheric degradation phenomenon according to atmosphere light can adopt following formula to describe:

I(x,y)=t(x,y)J(x,y)+(1-t(x,y))A

Wherein, I (x, y) Atmospheric Degraded Image for producing, t (x, y) be the propagation in atmosphere function, A is the atmosphere light intensity, J (x, y) style of representative image, above-mentioned formula table understands the origin cause of formation of atmospheric degradation phenomenon, the variation that has comprised picture contrast and color, purpose of the present invention is utilized above-mentioned formula exactly, and known parameter or hypothesis solve and obtain J (x, y), but the unknown number number that this formula solves is greater than the equation number of listing, therefore, the present invention is usingd the dark primary priori of the propositions such as He Kaiming as the basic constraint condition solved, thereby effectively obtain the solution of model of the present invention,

Dark primary priori proposes on the Atmospheric Degraded Image statistical basis out of doors, a kind of dark object concept (Dark Object Subtraction of part, DOS), this theory is thought, in the overwhelming majority's non-sky regional area, a certain pixel always has at least one Color Channel and has extremely low numerical value, produces the very little image of a width area light intensity minimal value, and available following formula defines the image I of helping secretly generated _d(x, y):

$I_D(x,y)=\underset{c\&Element;\{r,g,b\}}{\min }\left(\underset{\mathrm{\&Omega;}}{\min }\left(I_c(x,y)\right)\right)$

Wherein, I _cthe dedicated tunnel that (x, y) is original image, Ω is adaptive image subblock, with (x, y) square area centered by, the self-adaptation piecemeal is a step important in the present invention, directly affects the post-processed quality, and the fast value of son hour, the details of propagation in atmosphere function is more, stereovision distinctness, but smoothness deficiency, can cause the local contrast distortion, otherwise, can effectively reduce the local contrast distortion, but the propagation in atmosphere functional image obtained is too single, cause details and the stereovision of image obvious not, can not effectively distinguish close shot and distant view image, for this situation, in order to reach the balance between distortion rate and details, choose the size of 4% maximal value of image row and column as sub-block, so also just avoided unified sub-block to the image of image when removing cloud and mist that vary in size, in addition, for the coarse phenomenon that prevents that depth jump from producing, avoid occurring gridiron pattern effect and halation vestige in result, need carry out medium filtering to helping image secretly,

Second step, the atmosphere light image is obtained: adopts the self-adaption two-dimensional empirical mode decomposition, after carrying out 5 decomposition, obtains original image corresponding atmosphere light image A (x, y), and carry out even local atmosphere light by low-pass filtering,

Bidimensional Empirical Mode Decomposition (Bidimensional Empirical Mode Decomposition, BEMD) be the new method of a Multi-scale model proposing on one dimension empirical mode decomposition basis, on time domain, the picture signal of a two dimension is decomposed in the two dimension of different scale and accumulate modular function (Bidimensional Intrinsic Mode Functions, BIMF) and the residue trend map picture (residue), they are the figure layers that comprise different frequency, it is a kind of frequency analysis method in the time domain scope, image through self-adaptation BEMD of the present invention, decompose the BIMF image and the trend map picture that obtain a plurality of different scales, the different frequency characteristic that they are comprising image, wherein: the BIMF image is comprising the HFS in image, embodying the detailed information in image, when its yardstick comprised reaches 5, its contained detailed information reaches 99.99%, remaining trend map picture is corresponding the atmosphere illumination of image, estimated value that can be using the trend map picture that obtains as the atmosphere light image,

While carrying out in this way atmosphere light image estimation, because self-adaptation BEMD self levies the apolegamy yardstick according to image is special, make it the extraction of details is tending towards to maximum, just can get rid of completely the detailed information in image, i.e. high fdrequency component; And picture noise also shows as high fdrequency component, reflect the atmosphere optical information so the luminance component obtained can be very pure, just can not be because of the error of atmosphere light image in follow-up processing, cause the large tracts of land bright areas such as the sky in image, inclined to one side white object, the water surface, produce the color distortion phenomenon when processing, as shown in Figure 2, concrete implementation step is as follows for the adaptive algorithm framework proposed in the present invention:

Step 1, by the local extremum point in the geodetic operator identification image in mathematical morphology;

Step 2, the Interpolation Property of Radial Basis Function method structure envelope surface that employing has been optimized, maximum point and minimum point are carried out respectively to interpolation arithmetic, obtain maximum point envelope surface and minimum point envelope surface after computing, two curved surface datas are averaging and obtain average envelope surface data, therefore, the upper lower envelope Solve problems in image space just changes into the discrete data point interpolation reconstruction problem of three-dimension curved surface, and solving the envelope plane can be expressed as:

$s\left(x\right)=c_0+c_{1}x+c_{2}y+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_i\mathrm{\&phi;}\left(\right||x-x_i|\left|\right)$

Wherein, s (x) is interpolation knot, c ₀, c ₁, c ₂, λ _ifor multinomial coefficient and radial basis function combination coefficient, ‖ ‖ is euclideam norm, and φ () is radial basis function, by solving N+3 unit large linear systems once, solves c ₀, c ₁, c ₂, λ _ivalue, and then substitution coordinate figure a little obtain whole interpolation curved surface, to the methods such as thin-plate, multiquadratics of having chosen of φ (), improve being chosen for of φ ():

$\mathrm{\&phi;}\left(R\right)=\left\{\begin{array}{cc}{(1-R)}^P\mathrm{PR},& R<1\\ 0,& R>1\end{array}\right.$

Wherein, R=‖ x-x _i‖ is Euclidean distance, and P is a constant coefficient;

Step 3, deduct the average envelope surface with former curved surface;

Step 4, judge whether to meet end condition, because the number of BIMF zero crossing can't be added up, thus can be by the constraint condition of BIMF during Bidimensional Empirical Mode Decomposition the stop condition as screening process, also can be by the efficient Cauchy-type condition of convergence:

$\mathrm{SD}=\underset{x}{\overset{M}{\mathrm{\&Sigma;}}}\underset{y}{\overset{N}{\mathrm{\&Sigma;}}}\frac{{|f_{k-1}(x,y)-f_k(x,y)|}^2}{f_{k-1}^2(x,y)}$

Wherein, f _k(x, y) is k layer self-adapting BEMD pixel value of (x, y) point on image while decomposing, and can to make SD be the number between 0.2 to 0.3;

Repeating step (one)～(three), until meet given end condition, obtain the 1st layer of BIMF image bimf ₁(x, y), deduct bimf with original image ₁(x, y) obtain the 1st layer of residual image residue1, to residue1 repeating step ()～(four), obtain successively N layer BIMF image and the N layer residual image of image, in said process, extreme point solves, the stop condition of planar interpolation and screening is the core of this algorithm, is generally to carry out 5 times to decompose, just can obtain pure atmosphere light image, so final result is expressed as:

$r_5(x,y)=f(x,y)-\underset{k=1}{\overset{5}{\mathrm{\&Sigma;}}}{\mathrm{bimf}}_k(x,y)$

Wherein, bimf _k(x, y) is k layer BIMF image, r ₅(x, y) be the trend map picture (atmosphere light image) after 5 layers of decomposition, some Atmospheric Degraded Images in reality, its atmosphere illumination is difficult to meet the local flatness of image, in order to overcome this shortcoming, improve the robustness of algorithm, can adopt the low frequency illuminance information in low-pass filtering smoothing processing atmosphere light image;

The 3rd step, the refinement of propagation in atmosphere function: adopt to help secretly most and process thought, by following formula:

$t_1(x,y)=1-w\underset{\mathrm{\&Omega;}}{\min }\left(\frac{I_D(x,y)}{A(x,y)}\right)$

Obtain rough propagation in atmosphere function t ₁, wherein Ω is adaptive image subblock, and w is the parameter that the validity in order to keep restoring rear image is introduced, and value between 0 to 1, get 0.95 in experiment usually;

The softmatting algorithm that adopts Levin to propose, to rough propagation in atmosphere function t ₁carry out the Laplacian Matrix correction, the energy equation below optimization carrys out refinement t ₁:

E(t)=t ^TLt+λ(t-t ₁) ^TU(t-t ₁)

Obtain the propagation in atmosphere function t of refinement, wherein λ is normalized parameter, and U is and the unit matrix of the equal size of image that L is Laplce's correction matrix, above-mentioned energy equation is carried out to sparse linear and mean:

(L+λU)t=λt ₁

Wherein, λ is a very little numerical value, is set as 10 in experiment ^-4; The a certain pixel of Laplce's correction matrix L (i, j) can be expressed as ^[12]:

$\underset{k|(i,j)\&Element;w_k}{\mathrm{\&Sigma;}}\{{\mathrm{\&delta;}}_{\mathrm{ij}}-\frac{1}{\left|w_k\right|}[1+{(I_i-{\mathrm{\&mu;}}_k)}^T{({\mathrm{\&Sigma;}}_k+\frac{\mathrm{\&epsiv;}}{\left|w_k\right|}{U}_3)}^{-1}(I_j-{\mathrm{\&mu;}}_k)\left]\right\}$

Wherein, δ _ijit is Kronecker function; μ _kand Σ _kwindow w _kthe average of middle pixel and covariance matrix; | w _k| be window w _kthe quantity of middle pixel; ε is normalized parameter; U ₃3 to take advantage of 3 unit matrix; When carrying out computing, image array is launched by column vector, be converted to the vector of one dimension, I _iand I _jbe designated as the value of i and j pixel under meaning in one-dimensional vector;

The 4th step, go to cloud and mist image color territory to proofread and correct: to the R of image, G, B three primary colors passage is gone respectively the cloud and mist computing:

$J_c(x,y)=\frac{I_c(x,y)-A(x,y)}{\max (t(x,y),t_0)}+A(x,y),c\&Element;\{r,g,b\}$

Wherein, I _cthe dedicated tunnel that (x, y) is original image, t ₀for the atmospheric dissipation function floor value of setting, usually get 0.1, yet the more difficult satisfied comfort level of human eye that returns to of the image of processing, therefore, the present invention has introduced the rgb color territory and proofreaied and correct: research and the statistics of human vision color-aware show, a large amount of human eye color-aware are image preferably, between its RGB three-component, very strong correlativity is arranged; The present invention is based on Imagine Macmillan storehouse and added up four groups of images that visual effect is good, obtained tri-component correlation matrixs of RGB, in its Image Fusion as desired value, between component related coefficient and statistical variance in Table 1,

Table 1 is through related coefficient between the color image components of statistics

In table, the related coefficient of any two component X and Y is defined as:

$r=\frac{{\mathrm{\&Sigma;}}_m{\mathrm{\&Sigma;}}_n(X_{\mathrm{mn}}-\stackrel{\&OverBar;}{X})(Y_{\mathrm{mn}}-\stackrel{\&OverBar;}{Y})}{\sqrt{{\mathrm{\&Sigma;}}_m{\mathrm{\&Sigma;}}_n{(X_{\mathrm{mn}}-\stackrel{\&OverBar;}{X})}^2{\mathrm{\&Sigma;}}_m{\mathrm{\&Sigma;}}_n{(Y_{\mathrm{mn}}-\stackrel{\&OverBar;}{Y})}^2}}$

Wherein:

for X, the Y average, obviously-1≤r≤1, so there is a kind of thinking in color correction: seek a linear transformation, the RGB component of image converted, make correlation matrix between its component meet table 1, so just can make to export color and meet human eye vision,

To arbitrary pixel m in the image J (x, y) that removes cloud and mist _xy=[m _r, m _g, m _b] ^t, carry out following rgb color territory and proofread and correct ^[13],

$\left\{\begin{array}{c}{n}_{\mathrm{xy}}=B^T{m}_{\mathrm{xy}}\\ C_n=B^T{C}_{m}B\end{array}\right.$

Wherein, n _xyfor proofreading and correct rear pixel vector, B is the color gamut correction matrix, C _m, C _nbe respectively and proofread and correct front and back image color interchannel covariance matrix, in order to solve color correction matrix B, can be first to positive definite matrix C _m, C _ncarry out the cholesky decomposition, obtain C _m=Q _m ^tq _m, C _n=Q _n ^tq _n, easily try to achieve B=Q _m ^-1q _n, establishing the correlation matrix that between component, related coefficient forms is R _n, easily know R _nwith C _nthere is following relation:

$C_n=\left[\begin{array}{ccc}{\mathrm{\&sigma;}}_{n1}& 0& 0\\ 0& {\mathrm{\&sigma;}}_{n2}& 0\\ 0& 0& {\mathrm{\&sigma;}}_{n3}\end{array}\right]R_n{\left[\begin{array}{ccc}{\mathrm{\&sigma;}}_{n1}& 0& 0\\ 0& {\mathrm{\&sigma;}}_{n2}& 0\\ 0& 0& {\mathrm{\&sigma;}}_{n3}\end{array}\right]}^T$

C for image to be corrected _mcan pass through

try to achieve (wherein

); Perhaps also can try to achieve according to relation in above formula, document passes through in [13] and establish σ _n1=σ _n2=σ _n3obtain the σ in above formula _nk(k=1,2,3), equate with power before and after guaranteeing computing;

In fact color need to keep original each component power proportions relation at timing, so just be unlikely to occur colour cast, institute is during in the hope of this standard deviation, mean allocation not in the present invention, but adopt each component proportion allocation criterion difference of original image, simultaneously because there is variance in related coefficient between target component, this is very important in experiment, in order more to meet statistical value, this step adopts average plus-minus variance as correlation matrix and R between two other component _nobtain respectively three width images after correction, be averaging afterwards and obtain last correcting image.

The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, all any modifications of doing within the spirit and principles in the present invention, be equal to and replace and improvement etc., within all should being included in protection scope of the present invention.

Claims (8)

1. a method of removing the Atmospheric Degraded Image cloud and mist, is characterized in that, the method for described removal Atmospheric Degraded Image cloud and mist comprises the following steps:

Help secretly image obtain and the adaptive decomposition of medium filtering, atmosphere light image obtains;

The refinement of propagation in atmosphere function;

The vision correction of color gamut.

2. the method for removal Atmospheric Degraded Image cloud and mist as claimed in claim 1, is characterized in that, helps image secretly and generate to adopt the image adaptive piecemeal to process thought, obtains the corresponding image I of helping secretly of original image helping secretly on priori theoretical foundation _d(x, y), and use median filter smoothness of image to process, available following formula defines the image I of helping secretly generated _d(x, y):

$I_D(x,y)=\underset{c\&Element;\{r,g,b\}}{\min }\left(\underset{\mathrm{\&Omega;}}{\min }\left(I_c(x,y)\right)\right)$

Wherein, I _cthe dedicated tunnel that (x, y) is original image; Ω is adaptive image subblock, is a square area centered by (x, y), for the coarse phenomenon that prevents that depth jump from producing, avoids occurring gridiron pattern effect and halation vestige in result, need carry out medium filtering to helping image secretly;

3. the method for removal Atmospheric Degraded Image cloud and mist as claimed in claim 1, it is characterized in that, the atmosphere light image is obtained and is adopted the self-adaption two-dimensional empirical mode decomposition, after carrying out 5 decomposition, obtain the corresponding atmosphere light image of original image A (x, y), and by low-pass filtering carry out even local atmosphere light.

4. the method for removal Atmospheric Degraded Image cloud and mist as claimed in claim 3, is characterized in that, the self-adaption two-dimensional empirical mode decomposition, and concrete implementation step is as follows:

Step 1, by the local extremum point in the geodetic operator identification image in mathematical morphology;

Step 2, the Interpolation Property of Radial Basis Function method structure envelope surface that employing has been optimized, maximum point and minimum point are carried out respectively to interpolation arithmetic, obtain maximum point envelope surface and minimum point envelope surface after computing, two curved surface datas are averaging and obtain average envelope surface data, therefore, the upper lower envelope Solve problems in image space just changes into the discrete data point interpolation reconstruction problem of three-dimension curved surface, and solving the envelope plane can be expressed as:

$s\left(x\right)=c_0+c_{1}x+c_{2}y+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_i\mathrm{\&phi;}\left(\right||x-x_i|\left|\right)$

Wherein, s (x) is interpolation knot, c ₀, c ₁, c ₂, λ _ifor multinomial coefficient and radial basis function combination coefficient, ‖ ‖ is euclideam norm, and φ () is radial basis function, by solving N+3 unit large linear systems once, solves c ₀, c ₁, c ₂, λ _ivalue, and then substitution coordinate figure a little obtain whole interpolation curved surface, to the methods such as thin-plate, multiquadratics of having chosen of φ (), improve being chosen for of φ ():

$\mathrm{\&phi;}\left(R\right)=\left\{\begin{array}{cc}{(1-R)}^P\mathrm{PR},& R<1\\ 0,& R>1\end{array}\right.$

Wherein, R=‖ x-x _i‖ is Euclidean distance, and P is a constant coefficient;

Step 3, deduct the average envelope surface with former curved surface;

Step 4, judge whether to meet end condition, because accumulateing the number of modular function zero crossing in two dimension can't add up, so can be by the constraint condition of accumulateing modular function in two dimension during Bidimensional Empirical Mode Decomposition the stop condition as screening process, also can be by the efficient Cauchy-type condition of convergence:

$\mathrm{SD}=\underset{x}{\overset{M}{\mathrm{\&Sigma;}}}\underset{y}{\overset{N}{\mathrm{\&Sigma;}}}\frac{{|f_{k-1}(x,y)-f_k(x,y)|}^2}{f_{k-1}^2(x,y)}$

Wherein, f _kthe pixel value of (x, y) point on image when (x, y) is k layer self-adapting Bidimensional Empirical Mode Decomposition, and can to make SD be the number between 0.2 to 0.3;

Repeating step (one)～(three), until meet given end condition, obtain accumulateing modular function image bimf in the 1st layer of two dimension ₁(x, y), deduct bimf with original image ₁(x, y) obtain the 1st layer of residual image residue1, to residue1 repeating step ()～(four), obtain successively accumulateing modular function image and N layer residual image in the N layer two dimension of image, in said process, extreme point solves, the stop condition of planar interpolation and screening is the core of this algorithm, is generally to carry out 5 times to decompose, just can obtain pure atmosphere light image, so final result is expressed as:

$r_5(x,y)=f(x,y)-\underset{k=1}{\overset{5}{\mathrm{\&Sigma;}}}{\mathrm{bimf}}_k(x,y)$

Wherein, bimf _k(x, y) accumulates the modular function image in k layer two dimension, r ₅(x, y) be the trend map picture after 5 layers of decomposition, some Atmospheric Degraded Images in reality, its atmosphere illumination is difficult to meet the local flatness of image, in order to overcome this shortcoming, improve the robustness of algorithm, can adopt the low frequency illuminance information in low-pass filtering smoothing processing atmosphere light image.

5. the method for removal Atmospheric Degraded Image cloud and mist as claimed in claim 1, is characterized in that, the refinement of propagation in atmosphere function adopts to help secretly most processes thought, by following formula:

$t_1(x,y)=1-w\underset{\mathrm{\&Omega;}}{\min }\left(\frac{I_D(x,y)}{A(x,y)}\right)$

Obtain rough propagation in atmosphere function t ₁, wherein Ω is adaptive image subblock, and w is the parameter that the validity in order to keep restoring rear image is introduced, and value between 0 to 1, get 0.95 in experiment usually;

Adopt the softmatting algorithm, to rough propagation in atmosphere function t ₁carry out the Laplacian Matrix correction, the energy equation below optimization carrys out refinement t ₁:

E(t)=t ^TLt+λ(t-t ₁) ^TU(t-t ₁)

Obtain the propagation in atmosphere function t of refinement, wherein λ is normalized parameter, and U is and the unit matrix of the equal size of image that L is Laplce's correction matrix, above-mentioned energy equation is carried out to sparse linear and mean:

(L+λU)t=λt ₁

Wherein, λ is a very little numerical value, is set as 10 in experiment ^-4; The a certain pixel of Laplce's correction matrix L (i, j) can be expressed as:

$\underset{k|(i,j)\&Element;w_k}{\mathrm{\&Sigma;}}\{{\mathrm{\&delta;}}_{\mathrm{ij}}-\frac{1}{\left|w_k\right|}[1+{(I_i-{\mathrm{\&mu;}}_k)}^T{({\mathrm{\&Sigma;}}_k+\frac{\mathrm{\&epsiv;}}{\left|w_k\right|}{U}_3)}^{-1}(I_j-{\mathrm{\&mu;}}_k)\left]\right\}$

Wherein, δ _ijit is Kronecker function; μ _kand Σ _kwindow w _kthe average of middle pixel and covariance matrix; | w _k| be window w _kthe quantity of middle pixel; ε is normalized parameter; U ₃3 to take advantage of 3 unit matrix; When carrying out computing, image array is launched by column vector, be converted to the vector of one dimension, I _iand I _jbe designated as the value of i and j pixel under meaning in one-dimensional vector;

6. the method for removal Atmospheric Degraded Image cloud and mist as claimed in claim 1, is characterized in that, go to cloud and mist image color territory to proofread and correct concrete operations to be, and to the R of image, G, B three primary colors passage is gone respectively the cloud and mist computing:

$J_c(x,y)=\frac{I_c(x,y)-A(x,y)}{\max (t(x,y),t_0)}+A(x,y),c\&Element;\{r,g,b\}$

Wherein, I _cthe dedicated tunnel that (x, y) is original image, t ₀for the atmospheric dissipation function floor value of setting, usually get 0.1, very strong correlativity is arranged between the RGB three-component, four groups of images that visual effect is good have been added up based on Imagine Macmillan storehouse, tri-component correlation matrixs of RGB have been obtained, in its Image Fusion, as desired value, the related coefficient of any two component X and Y is defined as:

$r=\frac{{\mathrm{\&Sigma;}}_m{\mathrm{\&Sigma;}}_n(X_{\mathrm{mn}}-\stackrel{\&OverBar;}{X})(Y_{\mathrm{mn}}-\stackrel{\&OverBar;}{Y})}{\sqrt{{\mathrm{\&Sigma;}}_m{\mathrm{\&Sigma;}}_n{(X_{\mathrm{mn}}-\stackrel{\&OverBar;}{X})}^2{\mathrm{\&Sigma;}}_m{\mathrm{\&Sigma;}}_n{(Y_{\mathrm{mn}}-\stackrel{\&OverBar;}{Y})}^2}}$

Wherein:

for X, the Y average, obviously-1≤r≤1, so there is a kind of thinking in color correction: seek a linear transformation, the RGB component of image is converted.

7. the method for removal Atmospheric Degraded Image cloud and mist as claimed in claim 6, is characterized in that, to arbitrary pixel m in the image J (x, y) that removes cloud and mist _xy=[m _r, m _g, m _b] ^t, carry out following rgb color territory and proofread and correct,

$\left\{\begin{array}{c}{n}_{\mathrm{xy}}=B^T{m}_{\mathrm{xy}}\\ C_n=B^T{C}_{m}B\end{array}\right.$

Wherein, n _xyfor proofreading and correct rear pixel vector, B is the color gamut correction matrix, C _m, C _nbe respectively and proofread and correct front and back image color interchannel covariance matrix, in order to solve color correction matrix B, can be first to positive definite matrix C _m, C _ncarry out the cholesky decomposition, obtain C _m=Q _m ^tq _m, C _n=Q _n ^tq _n, easily try to achieve B=Q _m ^-1q _n, establishing the correlation matrix that between component, related coefficient forms is R _n, easily know R _nwith C _nthere is following relation:

$C_n=\left[\begin{array}{ccc}{\mathrm{\&sigma;}}_{n1}& 0& 0\\ 0& {\mathrm{\&sigma;}}_{n2}& 0\\ 0& 0& {\mathrm{\&sigma;}}_{n3}\end{array}\right]R_n{\left[\begin{array}{ccc}{\mathrm{\&sigma;}}_{n1}& 0& 0\\ 0& {\mathrm{\&sigma;}}_{n2}& 0\\ 0& 0& {\mathrm{\&sigma;}}_{n3}\end{array}\right]}^T$

C for image to be corrected _mcan pass through

try to achieve, wherein

perhaps also can try to achieve according to relation in above formula;

8. the method for removal Atmospheric Degraded Image cloud and mist as claimed in claim 6, is characterized in that, goes to cloud and mist image color territory to proofread and correct and adopt average plus-minus variance as correlation matrix and R between two other component _nobtain respectively three width images after correction, be averaging afterwards and obtain last correcting image.

Images