CN105513025A - Improved rapid demisting method - Google Patents

Abstract

The invention discloses an improved rapid demisting method. The method is divided into three parts: the first part involves performing mean-value down-sampling processing on YUV data and converting the YUV data into small graphs in an RGB format; the second part involves performing maximum-value filtering on small graph transmissivity for the purpose of solving the problem of abrupt change edge blacking caused by down-sampling; and the third part involves performing up-sampling by use of bilinear interpolation for the block effect caused by the down-sampling so as to prevent the block effect. The method brought forward by the invention can process a 5-million-pixel video in a YUV format in real time on a PC. An effective improvement method is brought forward for effectively prevent a black-edge side effect caused by the down-sampling and the up-sampling. The up-sampling by use of the bilinear interpolation is brought forward for removing the block effect generated due to the down-sampling.

Description

A kind of rapid defogging method of improvement

Technical field

The present invention relates to the technical field of Digital Image Processing, intelligent transportation, be specifically related to a kind of rapid defogging method of improvement.

Background technology

In recent years, environmental pollution states was very serious, and especially the weather of haze gets more and more.Mist, haze are a kind of aerial particles that suspends, and its scattering process greatly reduces picture quality.The reduction of picture quality directly affects the function of the outdoor image such as knotmeter, drive recorder acquisition system.Not only reduce the visuality of image, more difficulty is caused to the carrying out of successive image Processing Algorithm (as Car license recognition, feature extraction, graphical analysis etc.).Image collecting device is nowadays in the majority with high-definition camera, the clear picture of collection, but the data volume of image is also comparatively large, deals with consuming time longer.Therefore invent and a kind ofly the method for Quick demisting can seem very important.

But existing mist elimination algorithm performance and function are conflict bodies, and effective algorithm complex is high, are difficult to reach real-time; Fireballing algorithm is comparatively simple, but result is undesirable, often causes the spinoffs such as the distortion of color.Therefore need to carry out a balance to performance and function two aspect, propose a kind of method, in performance with functionally there is good performance.

Existing image mist elimination disposal route has a lot, can be divided into two large classes generally: based on the method for image enhaucament and the method for physically based deformation model.The method of image enhaucament strengthens the image be degraded, and improves the quality of image.The method is comparatively simple, and processing speed is very fast, but treatment effect is undesirable, may cause the loss of image portion information, so that image fault.The method of physically based deformation model, this method to the scattering process of light by research air suspended particle, is set up atmospherical scattering model, is understood the Physical Mechanism of image degradation, and inversion restoration goes out without mist image, nowadays a lot of mist elimination algorithm is all that physically based deformation model proposes.Prior art one (HeK, SunJ, TangX.Singleimagehazeremovalusingdarkchannelprior.IEEETr ansactionsonPatternAnalysisandMachineIntelligence, 2011,33 (12): 2341-2353) by a large amount of without mist image statistics observation of characteristics, found the priori rule being named as dark primary priori.The method has extraordinary performance on treatment effect, opens a frontier of image mist elimination.But adopt soft stingy figure to carry out refinement transmissivity figure in literary composition, complexity is very high, length consuming time, author used again Steerable filter to replace the mode of soft stingy figure afterwards, and mist elimination effect is suitable, and processing speed but improves about 100 times.Even if but use Steerable filter to carry out mist elimination for HD video, want to realize real-time process, also have very large gap.Prior art two (TarelJP, HautiereN.Fastvisibilityrestorationfromasinglecolororgra ylevelimage.In:Proceedingsofthe12thIEEEInternationalConf erenceonComputerVision, 2009.Kyoto:IEEE, 2009.2201-2208) in, propose a kind of method of Quick demisting, use the mini-value filtering in two medium filtering replacement prior art one and Steerable filter, enormously simplify processing procedure, raise the efficiency.But the holding edge filter algorithm that medium filtering has not been, the sudden change of the regional area depth of field can produce halo effect.And the parameter in algorithm is more, cannot self-adaptative adjustment be realized, need manually to carry out testing and debugging, be restricted in actual applications.

Domestic also have a lot of research institution or colleges and universities in mist elimination, have good achievement.Prior art three (greasy weather processing system for video and method CN103347171A based on DSP) devises a set of disposal system for greasy weather video, comprise the design of hardware and software, in program optimization for some change parameter slowly adopt every time the method that upgrades reduce the processing time.But can the image that author finally only mentions for 432*283 carries out simulation process 3.622s consuming time, be not specifically shown in DSP platform and reach real-time for HD video.Prior art four (CN104240192A, denomination of invention: one is single image mist elimination algorithm fast) in by minimum value, image gradient and the dark figure in the RGB triple channel of image color space with specific condition Fast back-projection algorithm go out mist elimination model need transmission plot, instead of original dark primary to test soft stingy figure method in mist elimination algorithm and solve the step of transmission plot, and optimize the calculating of dark.The computing of original Large Scale Sparse matrix has been become the comparison to the different frame corresponding pixel points of a few width by this method, and operand reduces greatly, and in most of the cases can obtain the result desirable on an equal basis with former algorithm effect.

Summary of the invention

Object of the present invention is: 1) for the traffic video of 500W pixel yuv format, and the method that the present invention proposes can reach real-time process on PC; 2) for the problem of black borders that down-sampled and liter sampling bring, the method that the present invention proposes can well avoid this spinoff.3) for the down-sampled blocking effect brought, the present invention uses bilinear interpolation to rise sampling, can avoid blocking effect.

The technical solution used in the present invention is: a kind of rapid defogging method of improvement, and the method step is as follows:

Step 1), input the image of a frame yuv format;

Step 2), yuv data average is down-sampled, and is converted into rgb format;

The down-sampled process of average is carried out to yuv data, obtain down-sampled after the image of yuv format, and the image of yuv format is converted to the image of rgb format;

Step 3), ask for little figure transmissivity based on dark primary priori principle;

For step 2) in the little figure of RGB, ask for dark by mini-value filtering, estimation air light value, calculates the transmissivity estimated, finally carrying out Steerable filter to estimating transmissivity, trying to achieve the transmissivity become more meticulous;

Step 4), maximal value filtering;

The liter sampling carrying out bilinear interpolation after radius is the maximal value filtering process of 1 is again carried out, to avoid the phenomenon of black surround to little figure transmissivity;

Step 5), bilinear interpolation rise sampling;

Adopt bilinear interpolation to rise sampling, between several point, carry out linear interpolation, make the image change after liter sampling mild, blocking effect can be eliminated;

Step 6) directly revert to YUV image without mist;

Utilize step 5) the large figure of transmissivity after interpolation, carries out the reconstruct of mist elimination image, directly adopts yuv format to be reconstructed, eliminate the Time and place that middle YUV and RGB changes;

Step 7) input next frame YUV image.

The advantage of technical solution of the present invention and good effect are:

1) YUV average is down-sampled, and is converted into rgb format;

The down-sampled process of average is carried out to yuv data, the down-sampled information that to a certain degree can ensure original image of average, obtain down-sampled after the image of yuv format, and the image of yuv format is converted to the image of rgb format.Greatly can reduce that calculating brings like this consuming time and save a part of storage space.

2) solution of break edge problem of black borders;

Use down-sampled and rise sampling and can bring certain spinoff, especially for the billboard having color darker in large stretch of sky areas or vehicle, the place that the rear advertising board of process or vehicle body and sky have a common boundary will be made like this to occur the situation that mist elimination is excessive, the namely appearance of black surround.The present invention to the transmissivity in algorithm carry out radius be 1 maximal value filtering process effectively can avoid the phenomenon of this black surround.

3) down-sampled blocking effect is avoided;

In the present invention, carry out bilinear interpolation for the little figure of transmissivity and rise sampling processing, utilize the large figure of transmissivity obtained to carry out the reconstruct of image, finally can obtain good mist elimination effect, and there is no this spinoff of blocking effect.

Accompanying drawing explanation

Fig. 1 is that bilinear interpolation rises sampling schematic diagram;

Fig. 2 is the process flow diagram of improving one's methods;

Fig. 3 is mist elimination result.

Embodiment

The present invention is further illustrated below in conjunction with accompanying drawing and specific embodiment.

Technical scheme of the present invention is divided into three parts: Part I carries out the down-sampled process of average to yuv data, and convert the little figure of rgb format to; Part II is to solve the down-sampled problem bringing break edge blackening, carries out maximal value filtering to little figure transmissivity; Part III is for the down-sampled blocking effect brought, and adopts bilinear interpolation to rise sampling and avoids blocking effect.

1) YUV average is down-sampled, and is converted into rgb format

The down-sampled process of average is carried out to yuv data, obtain down-sampled after the image of yuv format, and the image of yuv format is converted to the image of rgb format.

2) solution of break edge problem of black borders

In order to speed up processing, employ down-sampled the liter with bilinear interpolation of average and sample, but can bring certain spinoff, especially break edge there will be the excessive situation of mist elimination, namely the appearance of black surround.

For this situation in the present invention, the liter sampling carrying out bilinear interpolation is again carried out to the transmissivity in algorithm, effectively can avoid the phenomenon of this black surround after radius is the maximal value filtering process of 1.

3) bilinear interpolation rises sampling

Rise sampling be divided into close on value rise sampling and bilinear interpolation rise sample bicubic interpolation liter sample.Close on value rise sampling be in a block using that nearest pixel point value as in this block value a little.It is best that bicubic interpolation rises sample effect, but algorithm complex is higher, is not suitable in the present invention using.Adopting bilinear interpolation to rise sampling in the present invention, is carry out linear interpolation between several point, and make the image change after liter sampling mild like this, blocking effect can be eliminated.The principle of bilinear interpolation is as shown in Figure 1:

1,2,3,4 pixel values representing little figure respectively in Fig. 1, other blank parts are exactly the pixel value of up-sampling interpolation.Appoint get in first piece a bit, ask for the value of this point, the value of this point be near its 4 pixel values (1 in corresponding diagram, 2,3,4 pixel values) a weighted mean.Weighting coefficient is with distance dependent system, and distance is nearer, and weighting coefficient is larger, and weighting coefficient far away is less.Computing formula is as follows:

dest＝a*value1+b*value2+c*value3+d*value4

Wherein dest represents pixel value to be inserted, and valua1, valua2, valua3, valua4 represent the value of four point of proximity respectively, and a, b, c, d represent weighting coefficient, and they can be asked for according to following formula:

$a=\frac{(ratio-r)*(ratio-c)}{{\mathrm{ratio}}^2}$

$b=\frac{r*(ratio-c)}{{\mathrm{ratio}}^2}$

$c=\frac{(ratio-r)*c}{{\mathrm{ratio}}^2}$

$d=\frac{r*c}{{\mathrm{ratio}}^2}$

In above formula, ratio represents the multiplying power of up-sampling, and (r, c) in above formula molecule represents the coordinate of pixel to be asked in ratio*ratio block.

As shown in Figure 2, embodiment of the present invention is exemplified below:

1, the little figure of YUV average down-sampled one-tenth RGB

The present invention is directed to the traffic video of high definition, such as the video of 500W size, every two field picture size is 2432*2048.First average is carried out for yuv data down-sampled, doubly down-sampled for n at this, the process of averaging is carried out respectively for Y-component and UV component.Ask the average in n*n window to save, directly convert the image of RGB to according to the transformational relation between YUV and RGB, conversion formula is as follows:

R＝Y+1.371*(V-128)

G＝Y-0.698*(V-128)-0.336*(U-128)

B＝Y+1.732*(U-128)

Wherein Y, U, V represent the value of YUV tri-components respectively, R, G, B represent conversion after the value of RGB tri-passages.

2, the method based on dark primary priori tries to achieve transmissivity

The present invention is the physical model based on there being mist image, has the physical model of mist image to be expressed as:

I(x)＝J(x)t(x)+A(1-t(x))

Wherein, I (x) is exactly existing image (treating the image of mist elimination), and J (x) is the image without mist that will recover, and A is global atmosphere light component, and t (x) is transmissivity.

2.1, dark is obtained

In the regional area of the non-sky of the overwhelming majority, certain some pixel always has at least one Color Channel and has very low value.In other words, the minimum value of this area light intensity is a very little number.First obtain the minimum value in each pixel RGB component, dark to ask for formula as follows:

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

J in formula ^crepresent each passage pixel value of coloured image, Ω (x) represents a window centered by pixel x.The theory of dark primary priori is pointed out: J ^dark(x) → 0.

2.2, air light value is asked for

According to the physical model having mist image, want to recover without mist image, prerequisite knows air light value A.The method asking for A value in the present invention is: from dark figure, get the pixel of front 0.01% according to the size of brightness; In these positions, from original have mist image I find corresponding point and average, as A value.

2.3, the transmissivity estimated is calculated

By the distortion having mist image, and in conjunction with the theory of dark primary priori, transmissivity can be derived expression formula, as follows:

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

Wherein transmissivity, I ^crepresent three-channel value, A ^crepresent air light value;

2.4, Steerable filter

Steerable filter process is carried out to the above-mentioned transmissivity estimated, as input figure, gray-scale map, as guiding figure, obtains the little figure of transmissivity become more meticulous after process.

3, maximal value filtering is carried out for the little figure of transmissivity

Mist elimination algorithm based on dark primary priori has such rule: partially white region needs the degree of mist elimination high, and corresponding transmissivity will be smaller; Darker region (as vehicle body or billboard) needs the degree of mist elimination low, so the transmissivity of correspondence is higher.If there is darker billboard or vehicle body in large stretch of sky areas, so just there will be the spinoff of the edge blackening that billboard or vehicle body and sky areas have a common boundary.If sky areas transmissivity t1=0.5, billboard area transmissivity t2=0.9, so can plug the number between 0.5 ~ 0.9 between t1 and t2.Because the formula of Image Reconstruction is: J=(I-A)/t+A, vehicle body or billboard pixel value I natively smaller, A is that air light value is (larger, general more than 200), if t is a little bit smaller a little, the value of J will be made to be less than zero, so there will be the situation of blackening.

In order to address this problem, in the present invention, the maximal value filtering process that radius is 1 be carried out to little figure transmissivity, the appearance of black surround can be avoided like this, certainly can bring slight halo effect like this, but this impact and problem of black borders being compared insignificant.

4, bilinear interpolation rises sampling

For the little figure of transmissivity that step 3 obtains, carry out the process that bilinear interpolation rises sampling.In this example, carry out 16 times of bilinear interpolations and rise sampling, that is between little figure transmissivity two points, insert 15 points, make through the transmissivity image of interpolation equal with the length and width of former figure.But should be noted that the process at edge when interpolation, preventing liter transmissivity after sampling and former figure can not one_to_one corresponding, and result there will be some edge effects.

5, without the reconstruct of mist image

The transmissivity of the large figure obtained by step 4, next just can carry out the reconstruct without mist image.Directly omit intermediate conversion process in the present invention, directly to yuv format have mist image be reconstructed into YUV without mist image.First calculate

$\left\{\begin{array}{c}{A}_y=0.114*A_b+0.587*A_g+0.299*A_r\\ A_u=0.5*A_b-0.332*A_g-0.168*A_r+128\\ A_v=-0.082*A_b-0.418*A_g+0.5*A_r+128\end{array}\right.$

A in above formula _b, A _g, A _rrepresent the air light value of B, G, R tri-passages respectively, A _y, A _u, A _vrepresent the air light value after corresponding to Y, U, V.

Due to the A in a two field picture _ba _ga _rbe identical, the every frame of calculating above only needs to calculate once.Then reconstruct is without the YUV image of mist:

$\left\{\begin{array}{c}{Y}^{\′}=(Y-A_y)/t\left(x\right)+A_y\\ U^{\′}=(U-A_u)/t\left(x\right)+A_u\\ V^{\′}=(V-A_v)/t\left(x\right)+A_v\end{array}\right.$

In above formula, Y, U, V represent the data of original image, and t (x) represents transmissivity, A _y, A _u, A _vrepresent correspond to Y, U, V after air light value, Y', U', V' represent reconstruct after view data.

Result is analyzed as follows:

The method that the present invention proposes is on PC, and the YUV image process for 2432x2048 size is consuming time at about 30ms, reaches the requirement of real-time.And there is good treatment effect, as shown in Figure 3, wherein Fig. 3 (a) to be former figure, Fig. 3 (b) be carry out maximal value filtering for little figure transmissivity before image after mist elimination, so there is problem of black borders at vehicle body edge, as shown in the circle in Fig. 3 (b).Fig. 3 (c) be carry out maximal value filtering for little figure transmissivity after reconstruct out without mist image.Obviously find out that after adding maximal value filtering for little figure transmissivity, the problem of black surround can not occur again, and mist elimination ability can not be affected.

Claims (1)

1. the rapid defogging method improved, it is characterized in that, the method step is as follows:

Step 1), input the image of a frame yuv format;

Step 2), yuv data average is down-sampled, and is converted into rgb format;

The down-sampled process of average is carried out to yuv data, obtain down-sampled after the image of yuv format, and the image of yuv format is converted to the image of rgb format;

Step 3), ask for little figure transmissivity based on dark primary priori principle;

For step 2) in the little figure of RGB, ask for dark by mini-value filtering, estimation air light value, calculates the transmissivity estimated, finally carrying out Steerable filter to estimating transmissivity, trying to achieve the transmissivity become more meticulous;

Step 4), maximal value filtering;

The liter sampling carrying out bilinear interpolation after radius is the maximal value filtering process of 1 is again carried out, to avoid the phenomenon of black surround to little figure transmissivity;

Step 5), bilinear interpolation rise sampling;

Adopt bilinear interpolation to rise sampling, between several point, carry out linear interpolation, make the image change after liter sampling mild, blocking effect can be eliminated;

Step 6) directly revert to YUV image without mist;

Utilize step 5) the large figure of transmissivity after interpolation, carries out the reconstruct of mist elimination image, directly adopts yuv format to be reconstructed, eliminate the Time and place that middle YUV and RGB changes;

Step 7) input next frame YUV image.