CN103136735A - Single image defogging method based on dual-scale dark channel - Google Patents

Abstract

The invention provides a single image defogging method based on a dual-scale dark channel. The method comprises the steps of enabling pixels of an original image to be normalized, normalizing various images with various bit widths through various coefficients, filtering each pixel in the original image through a minimum filter by using an array of the original image, obtaining a new minimum array, namely a high-scale dark channel image, building a high-scale dark channel model through the high-scale dark channel image, calculating and obtaining a high-scale dark channel, drawing a grey level statistical histogram for the high-scale dark channel, obtaining the brightest 0.1 percent of the pixels of the original image, obtaining an illumination coefficient by an arithmetic mean, obtaining the minimum of three color components on each pixel of the original image, obtaining a low-scale dark channel model, calculating a transfer function through the low-scale dark channel model, building a non-fog image restoration model through a vision model, the illumination coefficient and the transfer function, calculating three color channels in the same way according to the illumination coefficient and the transfer function, and obtaining the restored non-fog image in the restoration image.

Description

A kind of single image defogging method capable of helping secretly based on two yardsticks

Technical field

The invention belongs to Digital Image Processing and technical field of computer vision, relate to and a kind ofly utilize method that two yardsticks help secretly to help preferential method secretly and improve existing, realization is to original image mist elimination fast and effectively, that restores reaches good visual effect without the mist image, expanded the existing algorithm scope of application after improvement, strengthen visual effect, accelerated computing velocity.

Background technology

The outdoor scene image often descends because the impact that has mist or flue dust causes visual effect, even causes the unclear almost illegible of objective fuzzy.Therefore image mist elimination algorithm has a wide range of applications in Digital Image Processing and computer vision field, and it has not only strengthened visual effect, can also provide picture depth information for other subsequent algorithm.

At present image mist elimination algorithm commonly used utilizes multiple image or some extend informations to increase the data volume of processing more, such as the defogging method capable based on depth information needs picture depth information; Image mist elimination algorithm based on polarization needs light polarization information; Perhaps directly utilize a series of pictures collection to distinguish mist, cut apart.These methods need the larger calculated amount of more information, and are difficult for realizing.Be extracted on CVPR in 2009 based on helping preferential single width mist elimination algorithm secretly, its relatively simple process and good treatment effect make this algorithm obtain concern.But this algorithm still exists problem in processing: one, the Fitting Calculation is too complicated, needs larger processing consumption, is difficult for realizing; Two, when the large tracts of land white object is arranged in image, the result error is larger.Three, can not process high resolving power, high-bit width image.

Summary of the invention

For the deficiencies in the prior art, the objective of the invention is to propose a kind of method of utilizing two yardsticks to help secretly improves art methods, realize the image mist elimination of Simple fast, eliminate the error of art methods when processing the white object scene, and expansion algorithm is applied in high resolving power, high-bit width image.

For realizing such purpose, the technical scheme that the invention provides a kind of single image defogging method capable of helping secretly based on two yardsticks comprises following steps:

Step S1: the original image pixel is normalized to 0～1 interval, and carry out normalization for different bit wide images with different coefficients;

Step S2: utilize the original image matrix, to the mini-value filtering device filtering of each pixel in original image through one 15 * 15, obtain pixel minimum value in three Color Channels that this wave filter covers, and then 225 pixels are got minimum value again, obtaining a new minimum value matrix is namely that high yardstick is helped figure secretly;

Step S3: utilize high yardstick to help figure secretly and build high yardstick and help model secretly, calculate and obtain high yardstick and help secretly, high yardstick is helped secretly do the gray-scale statistical histogram, and calculating accumulative histogram, it is namely the sequence of completing pixel value, then 0.1% the brightest original image pixel of taking-up is done sums and is on average obtained the illumination coefficient, and the illumination coefficient is used for effectively getting rid of original image scene large tracts of land white object to the impact of computing illumination coefficient;

Step S4: three color components on each pixel of original image are got minimum value, obtain low yardstick and help model secretly, utilize low yardstick to help model secretly and calculate transfer function;

Step S5: utilize vision mode, illumination coefficient and transfer function to build without the mist image deblur model, according to the illumination coefficient with transfer function three Color Channels employing identical calculations, obtain the final recovery of restored image without the mist image; Described without the following expression of mist image deblur model I (x):

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

Wherein I for restore without the mist image, J is original image, A is the illumination coefficient, t (x) transfer function, x are the pixel values of current pixel point, t ₀Be a constant, its objective is to guarantee that divisor is non-vanishing, get in the present invention 0.1.

Beneficial effect of the present invention: a kind of single image mist elimination algorithm of helping secretly based on two yardsticks of the present invention, help priority algorithm analysis secretly to what prior art proposed, for the deficiency on the prior art algorithm, utilize the advantage that two yardsticks are helped secretly that existing algorithm is improved.Utilize high yardstick to help the computing illumination coefficient secretly, solved the error of calculation that causes due to adularescent object in image scene in existing algorithm.Utilize low yardstick to help secretly and calculate transfer function, avoided the process of match, increased considerably computing velocity, reduced calculating and storage consumption.Method of the present invention has reduced computation complexity to a great extent, has expanded simultaneously the usable range of algorithm, can carry out defogging to high resolving power high-bit width image, has strengthened the visual effect without the mist image of restoring.

Description of drawings

Fig. 1 is the inventive method one-piece construction.

Fig. 2 a-Fig. 2 d is that the inventive method is processed the image example, and wherein original image resolution is 450 * 600 pixels.

Fig. 3 a-Fig. 3 c is the Contrast on effect of algorithm of the present invention and existing algorithm process general pattern, and wherein original image resolution is 694 * 327 pixels.

Fig. 4 a-Fig. 4 b be algorithm application of the present invention at high resolving power high-bit width image example 1, wherein original image resolution is 1920 * 1080 * 3 * 16 bits.

Fig. 5 is algorithm application of the present invention original image in high resolving power high-bit width image example 2, and wherein original image resolution is 1920 * 1080 * 3 * 16 bits.

Fig. 6 be algorithm application of the present invention restore in high resolving power high-bit width image example 2 without the mist image, wherein original image resolution is 1920 * 1080 * 3 * 16 bits.

Fig. 7 a-Fig. 7 c is the present invention when processing has the scene picture that the large tracts of land white object exists and the effect contrast figure of existing algorithm process, and wherein original image resolution is 480 * 270 pixels.

Embodiment

Below in conjunction with accompanying drawing, embodiments of the invention are elaborated.The present embodiment is implemented under take technical solution of the present invention as prerequisite, provided detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

As shown in Figure 1, the algorithm flow of the present embodiment is divided into: pre-service, computing illumination coefficient, calculating transfer function, four parts of image restoration.

This example provides a kind of and has helped the mist elimination algorithm with respect to existing secretly based on helping the more simple more effective more practical two yardsticks of preferential algorithm secretly, specifically comprises the steps:

Step S1: pre-service.Image pixel is carried out normalization, show that the image bit wide is 8, the normalization divisor is 256; The high-bit width image is divided into 12 and 16 two kinds, and 12 normalization divisors are 4096; The normalization divisor of 16 is 65536.The pixel value of image is normalized in 0～1 scope, and carries out normalization for different bit wide images with different coefficients; Described coefficient is exactly normalized divisor, and following exactly 256,4096 and 65536 are divided into three kinds of situations:

1) for each pixel of general pattern divided by 256;

2) for 12 each pixel of bit high-bit width image divided by 4096;

3) for 16 each pixel of bit high-bit width image divided by 65536.

Step S2: utilize the original image matrix, to the mini-value filtering device filtering of each pixel in original image through one 15 * 15, obtain pixel minimum value in three Color Channels that this wave filter covers, and then 225 pixels are got minimum value again, obtaining a new minimum value matrix is namely that high yardstick is helped figure secretly;

Help the concentration that is used for the estimated image mist secretly, its proposition can not obtain the necessary depth information of image mist elimination in view of existing image mist elimination technical method in single image.Help secretly based on a kind of statistical observation model: help sequence rule secretly and think outdoor natural scene image without mist, after process is helped priority processing secretly, the brightness of most of pixel will approach zero, if help the pixel that exists a large amount of brightness higher in image secretly, these brightness should come from airborne fog or sky so.

Building high yardstick, help model secretly at first be to utilize to help J secretly ^DarkComputing formula as follows:

$J^{\mathrm{dark}}=\underset{c\∈\{r,g,b\}}{\min }\left(\underset{y\∈\mathrm{\Ω}\left(x\right)}{\min }\left(J^c\left(y\right)\right)\right)---(2-1)$

In formula, J represents that original image is the original image shown in Fig. 2 a, J ^cThe Color Channel of image J, i.e. R, G, three passages of B, c is Color Channel, r, g, b be interior three Color Channels of red, green, blue of representative color passage respectively, and x is the pixel value of current pixel point, y is the set of the field pixel of x, and Ω (x) is the neighborhood centered by x, J ^DarkBe helping secretly of J, help matrix secretly and be the matrix that the minimum value in neighborhood and three Color Channels of pixel in original image obtains, be used for estimating the concentration of mist in the present invention.Fig. 2 b illustrate recovery without the mist image;

According to formula (2-1), make that Ω (x) is 15 * 15 square x neighborhood in helping model secretly calculating high yardstick, help model J secretly therefore build high yardstick ^Dark-highComputing formula is:

$J^{\mathrm{dark}-\mathrm{high}}=\underset{c\∈\{r,g,b\}}{\min }\left(\underset{y\∈\mathrm{\Ω}\left(x\right)=15\×15\mathrm{pixels}}{\min }\left(J^c\left(y\right)\right)\right)---(2-2)$

J ^Dark-highBe that high yardstick is helped matrix, pixels representation unit secretly, high yardstick is helped design sketch such as Fig. 2 d secretly.

Step S3: computing illumination coefficient: utilize high yardstick to help figure secretly and build high yardstick and help model secretly, calculate and obtain high yardstick and help secretly, high yardstick is helped secretly do the gray-scale statistical histogram, and calculating accumulative histogram, it is namely the sequence of completing pixel value, then 0.1% the brightest original image pixel of taking-up is done sums and is on average obtained the illumination coefficient, and the illumination coefficient is used for effectively getting rid of original image scene large tracts of land white object to the impact of computing illumination coefficient.

In the conventional method, the point that generally believes gray-scale value maximum in natural image is the illumination coefficient often, directly original image is got the higher point of gray-scale value as the illumination coefficient.But due to the car of existence white in present image, white construction thing, and the impact of the factors such as light of car, the point of gray-scale value maximum is no longer the illumination coefficient, causes previous methods inaccurate when the computing illumination coefficient.Utilize high yardstick to help model secretly in the present invention and replace original image to the gray-scale value calculating of sorting, fine as the to avoid impact of human factor on natural image.

Utilize high yardstick to help model secretly and ask the brightest 0.1% pixel: help secretly for high yardstick and do statistic histogram and carry out normalization, then calculate accumulative histogram for normalized statistic histogram, the calculating pixel number is taken out 0.1% the highest sets of pixel values of gray-scale value.

The highest gray-scale pixels value set is averaging: the sets of pixel values that obtains is done sums on average, obtain the illumination coefficient A in algorithm of the present invention.

Step S4: three color components on each pixel of original image are got minimum value, obtain low yardstick and help model secretly, utilize low yardstick to help model secretly and calculate transfer function.Specifically be divided into following two computation processes:

Build low yardstick and help model secretly:

According to formula (2-1), make that Ω (x) be x in helping model secretly calculating low yardstick, i.e. minimum value operation is only completed in three Color Channels of each pixel, helps model J secretly therefore hang down yardstick ^Dark-lowComputing formula is:

$J^{\mathrm{dark}-\mathrm{low}}=\underset{c\∈\{r,g,b\}}{\min }\left(\underset{y\∈\mathrm{\Ω}\left(x\right)=x}{\min }\left(J^c\left(y\right)\right)\right)---(4-1)$

Low yardstick is helped design sketch such as Fig. 2 c secretly.

Transfer function be actually original image after removing the illumination coefficient each pixel value and recovery without the mist image in a ratio of each geometric position respective pixel value after removal illumination coefficient; It is that three color components on each pixel of original image are got minimum value that low yardstick is helped model secretly; Helping model secretly due to high yardstick is to get minimum value in a zone, so image texture information dropout, in the transfer function that we require to obtain, the ratio of each pixel value is not identical, and low yardstick to help model secretly be that each pixel is done separately minimum value, kept difference between pixel, namely texture information, transform function therefore adopt the low yardstick method of helping secretly to calculate; Set up low yardstick and help model secretly, obtain helping secretly and keep simultaneously image texture information.

Utilize low yardstick to help model secretly and obtain calculating transfer function t (x):

$t\left(x\right)=1-\underset{c}{\mathrm{\ω}\min }\left(\frac{J^{\mathrm{dark}-\mathrm{low}}\left(y\right)}{A^c}\right)---(4-2)$

A in formula ^cBe the expression of illumination coefficient in three Color Channels, because three Color Channels adopt same operation, so its implication is identical with A, illumination coefficient A is by obtaining in step S3, ω _cImplication be a control coefrficient, be used for to control the degree of mist elimination process mist elimination.In fact, mist to a certain degree can provide the depth information of image to human eye, allows the people can distinguish nearby scenery and scenery at a distance from the image of two dimension, so if the visual effect that mist is removed the image that obtains fully is not fine.So ω in the present invention _cBe used for the mist of remaining micro-concentrations, be taken as 0.95.

Step S5: utilize vision mode, illumination coefficient and transfer function to build without mist image deblur model I (x), adopt identical calculations according to illumination coefficient A and transfer function t (x) at three Color Channels, obtain the final recovery of restored image without the mist image;

Image restoration.In the present invention, the mist elimination process is based on the following expression of a vision mode J (x):

J(x)＝I(x)t(x)+A(1-t(x)) (5-1)

In formula I for restore without the mist image, J is original image, algorithm of the present invention is to help method secretly by two yardsticks, utilizes high yardstick to help model computing illumination coefficient A and low yardstick secretly and helps model secretly and calculate transfer function t (x), finally calculate recovery without the mist image I.

Derive according to formula (4-1), the computing formula without the mist image I that obtains restoring:

$I\left(x\right)=\frac{J\left(x\right)-A}{\max (t\left(x\right),t_0)}+A---(5-2)$

Wherein I for restore without the mist image, J is original image, A is the illumination coefficient, t (x) transfer function, x are the pixel values of current pixel point, t ₀Be a constant, its objective is to guarantee that divisor is non-vanishing, get in the present invention 0.1.

Calculate simply owing to utilizing two yardstick methods of helping secretly to carry out the image mist elimination in the present invention, solve the heavy burden of original method on calculating, make based on the single image mist elimination algorithm of helping secretly and can expand to high resolving power high-bit width image.

Original image shown in Fig. 3 a; The result of prior art algorithm process shown in Fig. 3 b; For the image examples in the prior art algorithm in the present invention algorithm carry out the algorithm process result to such as Fig. 3 c, visual effect is basic identical.

As Fig. 4 a and Fig. 4 b, the mist elimination effect that algorithm examples in the present invention is applied to high resolving power high-bit width image is shown, wherein Fig. 4 a illustrates original image, and Fig. 4 b illustrates algorithm process result in the present invention.

Original image shown in Figure 5 and recovery shown in Figure 6 without the mist image, can find out that algorithm of the present invention can process the mist elimination problem of high resolving power high-bit width image well.

Algorithm process when having the large tracts of land white object in image is to such as Fig. 7 a illustrates original image; Fig. 7 b illustrates former algorithm process effect; Fig. 7 c the inventive method treatment effect, detail are seen in figure can find out that the inventive method treatment effect obviously is better than having algorithm aspect grain details in square frame.The inventive method makes arithmetic speed obtain promoting significantly, and concrete data see Table 1:

Table 1 algorithm of the present invention and the contrast of existing algorithm arithmetic speed

The above; only be the embodiment in the present invention, but protection scope of the present invention is not limited to this, anyly is familiar with the people of this technology in the disclosed technical scope of the present invention; can understand conversion or the replacement expected, all should be encompassed in of the present invention comprise scope within.

Claims (2)

1. single image defogging method capable of helping secretly based on two yardsticks, its feature comprises that concrete steps are as follows:

Step S1: the original image pixel is normalized to 0～1 interval, and carry out normalization for different bit wide images with different coefficients;

Step S2: utilize the original image matrix, to the mini-value filtering device filtering of each pixel in original image through one 15 * 15, obtain pixel minimum value in three Color Channels that this wave filter covers, and then 225 pixels are got minimum value again, obtaining a new minimum value matrix is namely that high yardstick is helped figure secretly;

Step S3: utilize high yardstick to help figure secretly and build high yardstick and help model secretly, calculate and obtain high yardstick and help secretly, high yardstick is helped secretly do the gray-scale statistical histogram, and calculating accumulative histogram, it is namely the sequence of completing pixel value, then 0.1% the brightest original image pixel of taking-up is done sums and is on average obtained the illumination coefficient, and the illumination coefficient is used for effectively getting rid of original image scene large tracts of land white object to the impact of computing illumination coefficient;

Step S4: three color components on each pixel of original image are got minimum value, obtain low yardstick and help model secretly, utilize low yardstick to help model secretly and calculate transfer function;

Step S5: utilize vision mode, illumination coefficient and transfer function to build without the mist image deblur model, according to the illumination coefficient with transfer function three Color Channels employing identical calculations, obtain the final recovery of restored image without the mist image; Described without the following expression of mist image deblur model I (x):

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

Wherein I for restore without the mist image, J is original image, A is the illumination coefficient, t (x) transfer function, x are the pixel values of current pixel point, t ₀Be a constant, its objective is to guarantee that divisor is non-vanishing, get in the present invention 0.1.

2. single image defogging method capable as claimed in claim 1, is characterized in that, described image pixel carries out normalization, shows that the image bit wide is 8, and the normalization divisor is 256; The high-bit width image is divided into 12 and 16 two kinds, and 12 normalization divisors are 4096; The normalization divisor of 16 is 65536.

Images