CN103150708A - Image quick defogging optimized method based on black channel - Google Patents

Abstract

Provided is an image quick defogging optimized method based on a black channel. Edges of an image and non-edge areas of the image are processed through different templates to obtain a transmission picture, and a precise atmosphere light value is obtained by the fact that a sky area or an area with thickest fog is divided. The method greatly reduces consumption, improves the image recovery speed, and improves image recovery effects.

Description

Image Quick demisting optimization method based on black channel

Technical field

The present invention relates to image processing techniques, specifically a kind of image Quick demisting optimization method based on black channel.

Background technology

At present, in existing image defogging method capable, utilize the differentiated treatment mist elimination of multiple image under same scene different shooting angles or different weather condition, can obtain certain effect, but in actual applications, be difficult to obtain multiple image in the short time.Utilize the geometric model of known scene, or utilize artificial input depth information to limit too the practicality of method.

In recent years, utilize the single image mist elimination to obtain very much progress.Tan(R Tan.Visibility in bad weather from a single image[A] .Proc of IEEE CVPR[C] .Anchorage, Alaska:IEEE Computer Society, 2008:1-8.) the image ratio Misty Image observed under fine day has higher contrast, and the hypothesis neighborhood territory pixel has the identical dough softening, utilize and maximize the color contrast that the local contrast method is come Recovery image, obtained better effects on some scene image, but the method does not also meet real physical model, the easy distortion of the result that obtains.Fattal(R Fattal.Single image dehazing[A] .Proc of SIGGRAPH[C] .Los Angeles:ACM Transactions on Graphics, 2008,27 (3): 1-9.) hypothesis shade and transmissivity are incoherent, utilize the ICA model to estimate transmissivity, thereby obtain the mist elimination result, but the method can not well be processed the thick fog image.He(K He, J Sun, X O Tang.Single image haze removal using dark channel prior[J] .IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011, 33:1-13.) etc. the people proposed to help secretly sequence rule (statistics shows that most outdoor photos meet this rule), utilize this rule mist elimination, can obtain good effect to most of Outdoor Scene mist images, but keep off in 0 zone at some black channel and deviation can occur, adopt same atmosphere light value to cause restoration result regional area cross-color to entire image, and the soft matting algorithm of optimizing transmission plot can consume plenty of time and internal memory.

Summary of the invention

The present invention provides a kind of image Quick demisting optimization method based on black channel in order to overcome above-mentioned the deficiencies in the prior art, has promoted image optimization speed, has improved recovery effects.

Technical solution of the present invention is as follows:

A kind of image Quick demisting optimization method based on black channel is characterized in that the method comprises the following steps:

The first step is inputted pending image;

Second step, select day dummy section or mist the denseest zone:

To containing the image of day dummy section, choose a day dummy section, then choose atmosphere light A in the zone on high;

To not containing the image of day dummy section, choose mist the denseest zone, then choose atmosphere light A in mist the denseest zone;

The 3rd step, the marginal information of extraction image, after the fringe region expansion, edge zone and non-fringe region adopt respectively different templates to process the black channel that obtains entire image;

The 4th step, the transmissivity of computed image pixel, formula is as follows:

t(x)＝e ^-βd(x) (4)

Wherein, β is the atmosphere light-scattering coefficient, and d (x) is depth map;

The 5th step obtained according to second step the transmissivity that atmosphere light A and the 4th step obtain, and calculated the image after recovering, and formula is as follows:

$J\left(x\right)=\frac{I\left(x\right)-A}{\max (t\left(x\right),t_0)}+A---\left(2\right)$

Wherein, t ₀Be that minimum transmittance restriction is processed, result after causing recovering becomes noise spot for the transmissivity t (x) that prevents some pixel is too little, and I (x) is original image,

Described day dummy section or the mist the denseest zone chosen is specifically as the black channel I of a certain pixel x _min(x)＞T _vAnd the edge pixel number in this neighborhood of pixels and the ratio N of total edge pixel count _Edge(x)＜T _p, should the zone be day dummy section or mist the denseest zone, arrange day dummy section or mist the minimum value of the black channel pixel value in dense zone be T _v, neighborhood of pixels inward flange pixel count and total edge pixel count ratio N in day dummy section or mist the denseest zone is set _EdgeMaximal value be T _p:

T _v＝0.9I _{min_max}

$T_p=\frac{N_{\min \_\mathrm{edge}}}{0.9}$

Wherein, I _{Min_max}Be the max pixel value of black channel image, N _{Min_edge}Minimum value for neighborhood inward flange pixel count and the total edge pixel count ratio of all pixels.

Described edge zone is adopted different templates to process with non-fringe region and referred to: edge zone employing 3 * 3 little templates are processed, adopt 15 * 15 large forms to process to non-fringe region, then the black channel of fringe region and the black channel of non-fringe region are integrated the black channel that obtains entire image.

Principle of the present invention is as follows:

The basic model of Misty Image is:

I(x)＝J(x)t(x)+A(1-t(x)) (1)

In formula (1), J (x) is the radiation information from target, the transmissivity of t (x) expression target emanation information, J (x) t (x) is the direct decay of target emanation information, the electromagnetic wave information of having described from target arrives the energy of sensor through transmission medium; A is the atmosphere light component, and A (1-t (x)) represents that atmosphere light arrives the energy of sensor after overdamping.

If atmosphere light A and transmission plot t are known, just can utilize formula (2) to carry out the image mist elimination and process.

$J\left(x\right)=\frac{I\left(x\right)-A}{\max (t\left(x\right),t_0)}+A---\left(2\right)$

T wherein ₀Be that minimum transmittance restriction is processed, result after causing recovering becomes noise spot for the transmissivity t (x) that prevents some pixel is too little.Adopt t in the He algorithm ₀Be 0.1.

In the He algorithm, when atmosphere light A is known, can be in the hope of the transmissivity of a certain pixel, namely according to formula (3)

$t\left(x\right)=1-\underset{y\∈\mathrm{\Ω}\left(x\right)}{\mathrm{\ω}\min }\left(\underset{c}{\min }\left(\frac{I^{c\left(y\right)}}{A^c}\right)\right)---\left(3\right)$

Wherein, I ^c(y) refer to the Color Channel of a certain pixel y,

Can regard as

Black channel (the black channel J of a certain pixel x ^Dark(x) refer to interior all pixels of neighborhood Ω (x) of this pixel x

Minimum value, y is a certain pixel in this neighborhood, J ^cThe a certain Color Channel of pixel y, namely

A is known atmosphere light, and ω is the constant parameter that arranges in order to prevent thorough mist elimination from causing distortion, and it is 0.95 that ω is set in the He algorithm.

Utilize formula (3) when finding the solution, need in the neighborhood Ω (x) of certain pixel of hypothesis, its transmissivity all equates, and in reality, pixel might not equate with transmissivity between pixel, the transmission plot of therefore trying to achieve and actual certain deviation arranged, the transmission plot that causes obtaining are the shapes of one one.Block size is decided by the neighborhood template Ω size of selecting, and selects neighborhood Ω larger, and this deviation also can be larger.In the algorithm of He, it is 15 * 15 template that entire image is all adopted the Ω size, although template is larger, and J ^Dark(x) → 0 degree is larger, but equally also can cause the boundary information out of true, form serious milky spot phenomenon, and template is less, and transmission plot is more accurate.Adopt in addition soft matting algorithm that it is optimized after obtaining coarse transmission plot, picture for 400 * 600 sizes, the size of the Matting Laplacian matrix L in soft matting algorithm is 240000 * 240000, this is a very huge matrix, and its calculated amount is also just well imagined.

For need to consume in the He algorithm a large amount of internal memories and computing time this deficiency, the present invention carries out following modification:

At first farthest reduce the transmission plot deviation by changing neighborhood Ω size, improve accuracy.Because the image after recovering is that fringe region difference is apparent in view at the scene depth Sudden change region, namely maximum in edge's transmission plot error, and it is all very little in other most of area differentiation, therefore fringe region need to keep detailed information, should adopt little formwork calculation black channel, and in non-edge, brightness of image changes milder, the transmission plot error is less, and is less demanding to details, should adopt large form to calculate black channel.The black channel J of entire image ^DarkBe the black channel J of edge ^Dark _EdgeBlack channel J with non-edge ^Dark _NonedgeSum.

Secondly, when transmission plot is optimized, do not adopt soft matting algorithm, can greatly improve speed like this, save internal memory and consumption, shorten computing time.Because black channel itself has just represented depth information roughly: for black matrix, do not reflect extraneous light, if be not subjected to the impact of atmosphere light, its brightness is 0, is subject to the impact of atmosphere light, and black matrix brightness meeting increases; Within the specific limits, mist is denseer, and the particulate of day aerial floating dust is more, and the atmosphere light scattering is more serious, and the impact that black matrix is subject to atmosphere light is larger, and brightness is also just larger; Black matrix from sensor more away from, namely d is larger, mist is just denseer, its brightness is also just larger; Therefore under the greasy weather condition, black matrix is relevant from the gamma correction apart from d and black matrix of sensor, and namely the degree of depth is relevant with gamma correction; Black channel is the black matrix model that utilizes, when not being subjected to the atmosphere influence of light under fine day, be black, under the greasy weather condition, be subject to the impact of atmosphere light, brightness increases with distance, depth information so correspondingly is provided, the larger explanation mist of the brightness of black channel is denseer, object from the distance of sensor more away from, namely d is larger.Utilize the relation between black channel and the degree of depth, the communication theory formula (4) according to light in atmospheric medium calculates transmissivity:

t(x)＝e ^-βd(x) (4)

Wherein, β is the atmosphere light-scattering coefficient, and its scattering coefficient to all visible lights is the same, in uniform dielectric, can think that β is steady state value in addition, the depth map of d (x) for trying to achieve.

In addition, in the He algorithm, with cloudy day day dummy section or mist the max pixel value in dense zone as atmosphere light A, determine cloudy day sky dummy section or the mist the denseest zone of former Misty Image by seeking in black channel front 0.1% pixel maximum region.The method is for the very large Misty Image of sky dummy section or contain the less Misty Image of white object and can obtain atmosphere light value accurately, but for containing the bulk white object, its pixel value will cause certain error greater than the image of cloudy sky pixel value, because black channel can't filter the white object larger than template, for this deficiency, the present invention has carried out following modification, thereby can better search out day dummy section or mist the denseest zone.

For sky dummy section or mist the denseest zone, because the impact that is subject to atmosphere light is comparatively serious, its pixel value I _skyMust be relatively large in whole black channel image, establish the black channel image I _minMax pixel value be I _{Min_max}, I _sky∈ λ I _{Min_max}, λ ∈ [0.9,1]; Because this part zone is relatively level and smooth, details is relatively less in addition, the edge pixel number in neighborhood of pixels and the ratio N of total edge pixel count _EdgeLess, namely

λ ∈ [0.9,1], wherein, N _{Min_edge}All pixel N _EdgeMinimum value.Satisfy simultaneously these two conditions and be only in Misty Image day dummy section or mist the denseest zone.Introduced similar approach in document (Jing Yu, Qingmin Liao.Fast single image fog removal using edge-preserving smoothing.IEEE ICASS P, 2011:1245-1248.), but threshold value T _pValue be fixed value 0.001, do not have robustness.Here, we enlarge I _skyAnd N _EdgeValue, can enlarge so the scope in dense zone of day dummy section or mist, therefore get black channel pixel value I _minCount ratio N with edge pixel _EdgeThreshold value T _v, T _pBe respectively:

T _v＝0.9I _{min_max} (5)

$T_p=\frac{N_{\min \_\mathrm{edge}}}{0.9}---\left(6\right)$

Only has the I of satisfying _min(x)＞T _vAnd N _Edge(x)＜T _pPixel just be considered to day dummy section or mist the denseest zone;

Compared with prior art, the invention has the beneficial effects as follows:

1, because the present invention can choose day dummy section or mist the denseest zone more accurately, thereby search out atmosphere light value more accurately, make the result after recovery more accurate.

2, the present invention has improved the Optimizing Flow of transmission plot, the restored image that utilizes the restored image after optimizing more not optimize is apparent in view this principle of fringe region difference at the scene depth Sudden change region, ask for the marginal information of former Misty Image, fringe region and non-fringe region are asked for respectively black channel under its different templates, and according to the relation of black channel and the degree of depth, the transmission plot after being improved.Simultaneously, greatly reduced memory consumption, optimal speed can promote 10 to 15 times.And exponentially curved line relation between the transmission plot that the present invention finds the solution and depth information, the transmissivity that makes the object of infinite point obtain is more accurate, thereby has greatly improved the recovery effects of day dummy section.

Description of drawings

Fig. 1 (a) is the Misty Image that contains the bulk white object; Fig. 1 (b) is the black channel image; Fig. 1 (c) is the atmosphere light value region that the He algorithm obtains; Fig. 1 (d) is the atmosphere light value region that the present invention obtains.Fig. 1 (e) is that the He algorithm is not optimized transmission plot t'; Fig. 1 (f) is He algorithm optimization transmission plot t; Fig. 1 (g) uses mist elimination that t' calculates J' as a result; Fig. 1 (h) uses mist elimination that t calculates J as a result; Fig. 1 (i) is the transmission plot that obtains after the present invention only improves atmosphere light; Fig. 1 (j) is that the present invention improves atmosphere light and is optimized transmission plot afterwards; Fig. 1 (k) is the restoration result after the present invention only improves atmosphere light; Fig. 1 (l) is that the present invention improves atmosphere light and optimizes transmission plot mist elimination result afterwards.

Fig. 2 (a) is the Misty Image that another width contains the bulk white object; Fig. 2 (b) obtains the brightness maximum region in black channel image and He algorithm; Fig. 2 (c) is the atmosphere light value region that the He algorithm obtains; Fig. 2 (d) is the atmosphere light value region that the present invention obtains.Fig. 2 (e) is that the He algorithm is not optimized transmission plot t'; Fig. 2 (f) is He algorithm optimization transmission plot t; Fig. 2 (g) uses mist elimination that t' calculates J' as a result; Fig. 2 (h) uses mist elimination that t calculates J as a result; Fig. 2 (i) is the transmission plot that obtains after the present invention only improves atmosphere light; Fig. 2 (j) is that the present invention improves atmosphere light and is optimized transmission plot afterwards; Fig. 2 (k) is the restoration result after the present invention only improves atmosphere light; Fig. 2 (l) is that the present invention improves atmosphere light and optimizes transmission plot mist elimination result afterwards.

Fig. 3 (a) is a common Misty Image; Fig. 3 (b) is that the He algorithm is not optimized transmission plot t '; Fig. 3 (c) is He algorithm optimization transmission plot t; Fig. 3 (d) uses mist elimination that t' calculates J' as a result; Fig. 3 (e) uses mist elimination that t calculates J as a result; Fig. 3 (f) is the restoration result after the present invention only is optimized transmission plot; Fig. 3 (g) is the sky dummy section (representing with white portion) of trying to achieve in the present invention; Fig. 3 (h) is that the present invention improves atmosphere light and is optimized transmission plot afterwards; Fig. 3 (i) improves atmosphere light and optimizes transmission plot mist elimination result afterwards.

Fig. 4 (a) is a width greasy weather input picture; Fig. 4 (b) is that the He algorithm is not optimized transmission plot t '; Fig. 4 (c) is He algorithm optimization transmission plot t; Fig. 4 (d) uses mist elimination that t' calculates J' as a result; Fig. 4 (e) uses mist elimination that t calculates J as a result; Fig. 4 (f) is the scene depth region of variation | J-J ' |; Fig. 4 (g) is the sky dummy section (representing with white portion) of trying to achieve in the present invention; Fig. 4 (h) is the edge extracting that carries out with the sobel operator in the present invention; Fig. 4 (i) is the fringe region that in the present invention, expansion obtains; Fig. 4 (j) is the transmission plot that the present invention obtains; Fig. 4 (k) is the mist elimination result that the present invention obtains.

Fig. 5 is process flow diagram of the present invention.

Embodiment

Elaborate to of the present invention below in conjunction with drawings and Examples: the example that the present embodiment is implemented under take technical solution of the present invention as prerequisite provided detailed embodiment and process, but protection scope of the present invention should not be limited to following embodiment.

Please first consult Fig. 5, Fig. 5 is process flow diagram of the present invention, and as shown in the figure, the concrete implementation step of the present invention is as follows:

The first step is inputted pending image;

Second step, select day dummy section or mist the denseest zone:

When containing the image of day dummy section, choose a day dummy section, then choose atmosphere light A in the zone on high;

To not containing the image of day dummy section, choose mist the denseest zone, then choose atmosphere light A in mist the denseest zone;

The solution procedure of atmosphere light is specific as follows:

(1) ask the black channel I of original image I (x) _min(x), template size is 15 * 15;

(2) calculate the marginal information of original image gray-scale map, obtain bianry image;

(3) to the pixel value of each edge, calculate edge pixel number that its neighborhood contains and the ratio of total edge pixel count, obtain chart of percentage comparison as N _Edge(x);

(4) only has the I of satisfying _min(x)＞T _vAnd N _Edge(x)＜T _pPixel just be considered to day dummy section or mist the denseest zone;

(5) ask for the max pixel value of former figure as atmosphere light A value in the sky dummy section of trying to achieve or mist the denseest zone.

The 3rd step, the marginal information of extraction image, after the fringe region expansion, edge zone and non-fringe region adopt respectively different templates to process the black channel that obtains entire image;

The 4th goes on foot, and calculates the transmissivity of certain pixel, and formula is as follows:

t(x)＝e ^-βd(x) (4)

Wherein, β is the atmosphere light-scattering coefficient, and d (x) is depth map;

The optimization concrete steps of transmission plot are:

(1) extract the marginal information (as using the sobel operator) of image, obtain the most obvious regional I of graded in transmission plot _Edge

(2) expand this zone I _EdgeWidth, the bianry image after being expanded

(edge calculation information accounts for the ratio of total pixel, if ratio large (as greater than 50%), namely marginal information is abundanter, can carry out the small neighbourhood expansion, as expansion neighborhood neighbors=8; If ratio less (less than 50%) can suitably increase the expansion neighborhood, as neighbors=12);

(3) to this bianry image

Fringe region utilize little template (as 3 * 3) to ask black channel I ^Dark _Edge, non-fringe region utilizes large form (as 15 * 15) to ask black channel I ^Dark _Nonedge

(4) black channel of entire image is the black channel I of fringe region ^Dark _EdgeBlack channel I with non-fringe region ^Dark _NonedgeSum, namely

I ^dark＝I ^dark _edge+I ^dark _nonedge

(5) depth image d (x)=I ^Dark

(6) try to achieve corresponding transmission plot t (x) according to formula (4).

The 5th step obtained according to second step the transmissivity that atmosphere light A and the 4th step obtain, and calculated the image after recovering, and formula is as follows:

$J\left(x\right)=\frac{I\left(x\right)-A}{\max (t\left(x\right),t_0)}+A---\left(2\right)$

Wherein, t ₀Be that the minimum transmittance restriction is processed, t is set herein ₀=0.1, I (x) is original image.

The present invention has verified that the He algorithm finds the solution the incorrectness of atmosphere light to containing bulk white portion image, as shown in Fig. 1 (c), Fig. 2 (c).Obtain atmosphere light region accurately by the atmosphere light derivation algorithm in the present invention, as shown in Fig. 1 (d), Fig. 2 (d), and then try to achieve atmosphere light value A accurately.After trying to achieve transmission plot according to formula (3), substitution formula (2) restoration result after atmosphere light that is improved, as shown in Fig. 1 (k), Fig. 2 (k), result shows, to containing the image of bulk white portion, only revise atmosphere light recovery effects afterwards and be greatly improved than the He algorithm.

After the present invention has verified and only transmission plot has been optimized on the impact of recovery effects.As shown in Fig. 4 (f), utilize the transmission plot t' that do not optimize and the image of trying to achieve respectively after recovery with the transmission plot t after soft matting algorithm optimization is that fringe region difference is apparent in view at the scene depth Sudden change region.Utilize the present invention to obtain after transmission plot t and He algorithm obtain atmosphere light A, according to be improved restoration result after transmission plot of formula (2), as shown in Fig. 3 (f).(f) in comparison diagram 3 and (e) can find, the mist elimination of sky dummy section is processed, and the present invention is better than the He method.Because the He method is to utilize formula (3) to try to achieve the transmission plot of entire image, transmission plot and black channel are linear, if in image, atmosphere light A value is less than the max pixel value of image, the transmission plot of utilizing so formula (3) to try to achieve has negative value, and it dummy section and atmosphere light value are very approaching, cause the transmissivity that obtains very little, the easy distortion of image sky dummy section that recovers with formula (2).And we adopt formula (4), the curved relation of transmission plot and black channel, the transmission plot of calculating can guarantee be all on the occasion of, and be difficult for forming noise, because the present invention does not adopt soft matting algorithm, greatly promote computing velocity in addition, shortened computing time.

The present invention verified and improved simultaneously atmosphere light and optimize after transmission plot the impact of recovery effects, and step is as follows:

(1) ask the black channel I of Fig. 4 (a) I _min(x), adopting template size is 15 * 15;

(2) with the marginal information of its gray-scale map of sobel operator calculating chart 4 (a), obtain bianry image I _Edge, as Fig. 4 (h);

(3) to the pixel value of each edge, calculate edge pixel number that its neighborhood contains and the ratio of total edge pixel count, obtain chart of percentage comparison as N _Edge(x);

(4) ask for and satisfy I _min(x)＞T _vAnd N _Edge(x)＜T _pPixel region as the denseest zone of mist, as Fig. 4 (g);

(5) ask for the max pixel value of Fig. 4 (a) as atmosphere light A value in the mist the denseest zone of trying to achieve;

(6) to bianry image I _EdgeCarry out the neighborhood expansion, adopt neighbors=12, the bianry image after being expanded

, as Fig. 4 (i);

(7) to this bianry image Fringe region ask black channel I with 3 * 3 templates ^Dark _Edge, non-fringe region is asked black channel I with 15 * 15 template ^Dark _Nonedge, the black channel that obtains entire image is:

I ^dark＝I ^dark _edge+I ^dark _nonedge

(8) depth image d (x)=I ^Dark

(9) try to achieve corresponding transmission plot t (x) according to formula (4), as Fig. 4 (j);

(10) try to achieve atmosphere light A and transmission plot t (x) afterwards, utilize the result after formula (2) is tried to achieve recovery, as shown in Fig. 4 (k), t ₀Be 0.1.

Claims (3)

1. the image Quick demisting optimization method based on black channel, is characterized in that, comprises the following steps:

The first step is inputted pending image;

Second step, select day dummy section or mist the denseest zone:

When containing the image of day dummy section, choose a day dummy section, then choose atmosphere light A in the zone on high;

To not containing the image of day dummy section, choose mist the denseest zone, then choose atmosphere light A in mist the denseest zone;

The 3rd step, the marginal information of extraction image, after the fringe region expansion, edge zone and non-fringe region adopt respectively different templates to process the black channel that obtains entire image;

The 4th goes on foot, and calculates the transmissivity of certain pixel, and formula is as follows:

t(x)＝ ^e-βd(x) (4)

Wherein, β is the atmosphere light-scattering coefficient, and d (x) is depth map;

The 5th step obtained according to second step the transmissivity that atmosphere light A and the 4th step obtain, and calculated the image after recovering, and formula is as follows:

Wherein, t ₀The=0.1st, the minimum transmittance restriction is processed, and I (x) is original image.

2. the image Quick demisting optimization method based on black channel according to claim 1, is characterized in that, described day dummy section or the mist the denseest zone chosen is specifically as the black channel I of a certain pixel x _min(x)＞T _vAnd the edge pixel number in this neighborhood of pixels and the ratio N of total edge pixel count _Edge(x)＜T _p, should the zone be day dummy section or mist the denseest zone, arrange day dummy section or mist the minimum value of the black channel pixel value in dense zone be T _v, neighborhood of pixels inward flange pixel count and total edge pixel count ratio N in day dummy section or mist the denseest zone is set _EdgeMaximal value be T _p:

T _v＝0.9I _{min_max}

Wherein, I _{Min_max}Be the max pixel value of black channel image, N _{Min_edge}Minimum value for neighborhood inward flange pixel count and the total edge pixel count ratio of all pixels.

3. the image Quick demisting optimization method based on black channel according to claim 1, it is characterized in that, described edge zone is adopted different templates to process with non-fringe region and referred to: edge zone employing 3 * 3 little templates are processed, adopt 15 * 15 large forms to process to non-fringe region, then the black channel of fringe region and the black channel of non-fringe region are integrated the black channel that obtains entire image.

Images