CN107578386A - A kind of optimization defogging processing method of unmanned plane shooting image - Google Patents

Abstract

A kind of optimization defogging processing method of unmanned plane shooting image provided by the invention, based on dark channel prior defogging Processing Algorithm, for the characteristics of unmanned plane image data amount is big, map sheet is more, algorithm speed prioritization scheme is proposed, the dark for proposing to reduce data volume method using image down sampling and Algorithms T-cbmplexity is O (1) filters；Concurrently referring now to the uniform situation of cloud and mist in image, atmospheric transmissivity figure layer can be refined instead of guiding filtering using medium filtering, calculate the time so as to reduce, by speed-optimization, algorithm has well adapted to the characteristics of unmanned plane image data amount is big.For the higher uneven image of cloud and mist that is proposed for of unmanned plane image quality requirements during defogging, contrast stretching is carried out to transmissivity figure layer；Complete to carry out enhancing processing to image after defogging processing, by quality optimization, unmanned plane image defogging Processing Algorithm output image quality has obtained effective raising.

Description

A kind of optimization defogging processing method of unmanned plane shooting image

Technical field

The present invention relates to a kind of unmanned plane image defogging processing method, more particularly to a kind of optimization of unmanned plane shooting image Defogging processing method, belong to unmanned plane technical field of image processing.

Background technology

Unmanned plane is one kind of not manned vehicle, and it mainly passes through the presetting apparatus provided for oneself or wireless remotecontrol side Formula is manipulated.In the 1940s, unmanned plane occurs first, unmanned plane is mainly used in military use in World War II, to 20 The nineties in century, many countries were fully recognized that the features such as unmanned plane is simple in construction, use cost is cheap, no matter in economic development Or larger effect can be played in military affairs, and then started and a burst of using new and high technology development and developed into the height of unmanned plane Tide, nowadays as the continuation of the journey cycle of new material and the emergence unmanned plane of new aerofoil greatly increases, unmanned plane can use first The signal transacting that enters and the communication technology simultaneously can carry Miniature Sensor and obtain remote sensing images and complete task of taking photo by plane, and task of taking photo by plane uses Unmanned plane, which is used as to take photo by plane platform and carry airborne sensor, obtains data, and handles obtaining view data, according to certain Required precision makes image.Different airborne sensors is used according to task of taking photo by plane is different, can such as carry light optical camera, red Outer scanner, magnetic determining device, high resolution CCD digital camera, multi-spectral imager, laser scanner, synthetic aperture radar etc., institute The sensor of carrying is all high accuracy, small volume, the digital sensor of the larger excellent performance of amount of storage in light weight.

In actual task, because unmanned plane body is lighter, airborne weight is limited, and airborne sensor is mostly Miniature digital phase Machine, unmanned plane data map sheet is smaller compared with tradition takes photo by plane data, and amount of images is larger, and it is also relative to there is problem in image It is more, so unmanned plane image still has the characteristics of traditional Aerial Images data volume is big, due to amount of images increase simultaneously And more problem in image being present, processing unmanned plane image is more cumbersome, but unmanned plane during flying is highly relatively low, so the figure obtained As resolution ratio is higher, and there is very strong mobility, but also due to the relatively low reason of flying height, the figure of unmanned plane acquisition As being highly prone to ground or low latitude object and weather interference, particularly cloud and mist weather or ground smog are to unmanned plane image matter Amount is rather notable, so with reference to unmanned plane feature of image, realizes that the processing of unmanned plane image defogging is significant, to sum up The unmanned plane image has following a few point features：Image map sheet is more, more problems in image be present；Resolution ratio is higher, number It is larger according to measuring；Easily by ground or low latitude object and climatic effect.

The gradually popular in the prior art image defogging method of dark channel prior, base of this method based on dark channel prior On plinth, the thickness depth of cloud and mist and the radiation reduction amount of each pixel can be quantified based on this priori conditions, so as to The picture rich in detail of high quality is reduced, but image data amount is big, map sheet is more, image medium cloud because the processing of unmanned plane image defogging has The characteristics of mist situation complexity, the algorithm is applied to the processing of remote sensing images defogging and also needs to do some following improvement：

First, needing to improve calculating speed, reduce algorithm complex.Unmanned aerial vehicle remote sensing technology is with quick, motor-driven, flexible work Claim, particularly in disaster remote sensing fields, the defects of traditional emergency method can be made up well.In the usual calamity of disaster remote sensing fields Feelings are fast changing, it is desirable to which unmanned plane image processing algorithm is because of the characteristics of having quick and precisely, particularly as image defogging handles this The Preprocessing Algorithm of sample, although dark channel prior algorithm defog effect is notable, because algorithm complex is higher, particularly guide The Algorithms T-cbmplexity of filtering and dark algorithm is O (N^2), it is difficult to meet the processing of unmanned aerial vehicle remote sensing image big data quantity It is required that so need further to improve calculating speed.

Second, improve cloud and mist image processing effect pockety., can be sensitive because unmanned plane during flying is highly relatively low Detection ground object or low latitude climate change, and unmanned plane shooting image scene information is complicated, the thickness degree of cloud and mist It can change with the change of scene depth, the remote sensing compared with the image shot in life captured by unmanned plane under normal circumstances The distribution of image kind cloud and mist is not especially uniformly, if being carried out using dark channel prior defogging algorithm to cloud and mist image pockety Processing, image cloud and mist is distributed relatively thin part and obvious " cross and handle " just occurs, and cloud and mist is distributed thicker part and then occurred " undercuring " phenomenon that cloud and mist not yet removes completely, either " cross and handle " and " undercuring " can all largely effect on picture quality, In the remote sensing application higher to image quality requirements, Remote Sensing Image Quality may produce material impact to last decision-making.

Third, improve the defog effect of white portion similar to atmosphere light in picture.Dark channel prior principle is to big Amount outdoor image observed a kind of statistical result obtained afterwards, if in scene the radiation of light source intensity of some atural objects with Atmospheric radiation intensity approaches, then by these ground object areas are unsatisfactory for the precondition that algorithm assumed, such atural object area Domain can not obtain satisfied defogging result, and the status captured by unmanned aerial vehicle remote sensing image is complicated, such as in north of china in winter very Be possible to occur to be also possible in sheet of snowfield, farmland to occur substantial amounts of green house of vegetables and white roof in city or Cement road can all have an impact to dark channel prior defogging algorithm process result.

Fourth, reducing parameter setting, image defogging batch processing is realized.The image map sheet of unmanned plane shooting is smaller, amount of images It is more, if carrying out defogging processing to each image using dark channel prior defogging algorithm needs artificial setting quantity of parameters, such as Dark filter window size, greatly increase cost of labor and have a strong impact on operating efficiency, it is achieved that utilizing foggy image Adaptive defogging parameter itself is particularly significant, and this is also to fully achieve dark channel prior method to handle unmanned plane image Bottleneck problem.

The content of the invention

In view of the shortcomings of the prior art and area for improvement, a kind of unmanned plane shooting image provided by the invention is excellent Change defogging processing method, based on dark channel prior defogging Processing Algorithm, for unmanned plane image data amount is big, more than map sheet Feature, propose algorithm speed prioritization scheme.For reducing two aspect problems of data volume and optimized algorithm time complexity, propose Data volume method and Algorithms T-cbmplexity are reduced using image down sampling to filter for O (1) dark；Concurrently referring now to figure The uniform situation of cloud and mist as in, can be refined using medium filtering instead of guiding filtering to atmospheric transmissivity figure layer, so as to Reduce and calculate the time.By speed-optimization, algorithm has well adapted to the characteristics of unmanned plane image data amount is big.Due to unmanned plane Aerial Images are mainly used in agricultural, geology, hydrological environment monitoring and military surveillance etc., so for image quality requirements Higher, the present invention proposes two kinds of prioritization schemes for this problem：It is right for the uneven image of cloud and mist during defogging Transmissivity figure layer carries out contrast stretching；Complete to carry out enhancing processing to image after defogging processing.By quality optimization, nobody Machine image defogging Processing Algorithm output image quality has obtained effective raising.

To reach above technique effect, the technical solution adopted in the present invention is as follows：

A kind of optimization defogging processing method of unmanned plane shooting image, based on dark channel prior defogging processing method, With reference to the quality of data after unmanned plane feature of image optimization calculating speed and defogging, the defogging side for being adapted to unmanned plane shooting image is obtained Method；

Dark channel prior defogging processing method obtains priori according to dark principle, utilizes foggy image degeneration physics The processing of model realization image defogging, is divided into the following steps according to foggy image degeneration physical model by image defogging processing procedure： Estimate atmosphere illumination intensity A, estimation atmospheric transmissivity, defogging processing；

Enter with reference to the optimization defogging algorithm of unmanned plane feature of image in terms of processing speed optimization and treatment effect optimization two Hand, processing speed optimization include following three aspects：First, the dark that Algorithms T-cbmplexity is O (1) filters, second, using Image down sampling reduces data volume, third, being refined atmospheric transmissivity figure layer using medium filtering；Treatment effect optimization includes following Two aspects：When enhancing atmospheric transmissivity figure layer contrast, second, image enhancement processing.

A kind of optimization defogging processing method of unmanned plane shooting image, further, dark principle is in big discharge observation The statistical theory drawn on the basis of clear picture：In most of clear pictures not comprising sky, each pixel is in r, g, b The gray value of at least one passage is very low on three Color Channels or even levels off to 0, the spoke in the regional area of some very little It is very low to penetrate the minimum value of intensity, foggy image compares that brightness is higher and transmissivity t is relatively low with clearly fog free images, so having The dark channel value of mist image is higher, and dark channel image gray value is also higher, and dark channel image depicts the thickness degree of cloud and mist, borrows Dark figure layer is helped to solve to obtain atmospheric transmissivity.

A kind of optimization defogging processing method of unmanned plane shooting image, further, estimation atmosphere illumination intensity A are looked for first Gray value highest and the pixel of image total pixel number amount 0.1% is accounted for into dark channel image, record its position, and record Its corresponding coordinated indexing in the picture, found according to the coordinated indexing of record in the foggy image of input corresponding Pixel, and these pixel gray level average values found in foggy image are calculated as atmosphere illumination intensity.

The optimization defogging processing method of a kind of unmanned plane shooting image, further, with picture when estimating atmospheric transmissivity Atmospheric transmissivity is considered as invariant in filter window centered on plain x, the spectral window centered on pixel x is represented using t (x) Intraoral atmospheric transmissivity, the minimum value in each Color Channel is calculated respectively, calculate the filter window centered on pixel x The minimum value of interior gray scale, then air projection ratio can directly be calculated.

The optimization defogging processing method of a kind of unmanned plane shooting image, further, according to foggy image degradation model, figure Recover output clearly fog free images J by inputting foggy image I as defogging is handled, pass through following formula

By the atmospheric transmissivity t (x) in the filter window centered on pixel x, foggy image I (x), atmosphere illumination intensity A, you can draw fog free images J (x).

A kind of optimization defogging processing method of unmanned plane shooting image, further, Algorithms T-cbmplexity are O's (1) O (1) in dark filtering method is using the algorithm complex of the mini-value filtering of bucket sort, wherein 1 is that filter window is big It is small, because two the cumulative of histogram are O (1) operations, so as to per an increased pixel, histogram need to only tire out Add a fixed number, so as to obtain the complexity of O (1), specifically, Algorithms T-cbmplexity is O (1) mini-value filtering Algorithm by forming in several steps：

The first step, initialize core histogram；If image there is n to arrange, it is necessary to safeguard n histogram, each histogram pair Each row in image, in processing procedure, the histogram of each row will be updated, these histograms include with certain The information of 2r+1 pixel centered on individual pixel, r are filter window radius, and original state is during each first pixel of row is The histogram information of 2r+1 pixel of the heart, in processing, these histograms are updated simultaneously with the movement of sliding block；

Second step, update core histogram；More new histogram is as sliding block moves, and the histograms of each row is subtracted most left The histogram information of side pixel, while increase the histogram information of the pixel of the rightmost side, this step is also O (1) operation；

3rd step, obtain minimum value；Minimum value is taken according to the core histogram of renewal；

All single pixel operations are all unrelated with window size, and the mini-value filtering of above-mentioned gray level image is extended And it is applied to coloured image, and first are carried out by mini-value filtering, compares take r afterwards for r, the figure layer of tri- wave bands of g, b respectively, g, b tri- Each pixel that individual wave band figure layer was entered after mini-value filtering, minimum value in three pixel values is taken as dark channel value, Realize that the dark that Algorithms T-cbmplexity unrelated with filter window is O (1) filters.

A kind of optimization defogging processing method of unmanned plane shooting image, further, data are reduced using image down sampling Amount is the thumbnail for the data volume generation input picture for reducing input picture using the mode of image down sampling, utilizes input picture Thumbnail calculates the thumbnail of dark channel image, is finally carried out the thumbnail of dark channel image using the method used on image Amplification, further to be calculated.

A kind of optimization defogging processing method of unmanned plane shooting image, it is further, saturating using medium filtering air of refining Penetrate rate figure layer and solve halation phenomenon, median filter protects image edge information while being smoothed to image.

A kind of optimization defogging processing method of unmanned plane shooting image, further, the figure layer contrast of enhancing atmospheric transmissivity Degree is that the dark channel image of foggy image is first asked for when asking for atmospheric transmissivity and is normalized to obtain image G (x), image G (x) is taken against atmospheric transmissivity figure layer is obtained, is used to adjust atmospheric transmissivity wherein adding parameter ∈, ∈'s takes It is [0.9,1.2] to be worth scope：

T (x)=1- ∈ G (x)

To the image after dark defogging algorithm process, atmospheric transmissivity figure layer contrast stretching is carried out.

A kind of optimization defogging processing method of unmanned plane shooting image, further, image enhancement processing method are included certainly Dynamic contrast and auto color；

Most bright in image and most dark pixel is mapped to white and black by auto contrast, and the higher region of gray scale is brighter, The relatively low region of gray scale is darker, and auto contrast realizes that step is as follows：

1st step, set for determining parameter a HighCut and LowCut for passage upper lower limit value；

2nd step, to r, the image of tri- passages of g, b, distribution statisticses grey level histogram；

3rd step, respectively to the image setting parameter of each passage, so that it is determined that upper lower limit value, wherein N and M are distributed as image It is wide and high；

4th step, compare the maximum and minimum value of three passage upper lower limit values, with the minimum value of lower limit and the maximum of the upper limit As being newly upper and lower bound, and calculate and insinuate table；

5th step, root tool insinuate table, r, the image of tri- passages of g, b are insinuated respectively；

Auto color is that the contrast of image is adjusted according to the actual grey of image, the shade in clip image and High light value, then the most bright and most dark pixel of the remainder in image is mapped to pure white or black, intermediate pixel weighs in proportion New distribution distribution.

Compared with prior art, the present invention is special with reference to unmanned aerial vehicle remote sensing image based on dark channel prior defogging method Point, be improved for weak point in dark channel prior defogging method, according to feature of image for algorithm calculating speed and Improve the aspect of algorithm output image quality two and propose a variety of prioritization schemes, the advantage of the invention is that：

1. improving the calculating speed of unmanned plane image defogging processing method, algorithm complex is reduced.The present invention provides A kind of unmanned plane shooting image optimization defogging processing method, by Algorithms T-cbmplexity be O (1) dark filter, The optimization of the methods of atmospheric transmissivity figure layer of being refined using image down sampling reduction data volume, using medium filtering so that nobody More quick and precisely, the Preprocessing Algorithm that particularly image defogging is handled is more rapidly and efficiently so that dark for machine image processing algorithm Not only defog effect is notable for channel prior algorithm, and algorithm complex is relatively low, meets unmanned aerial vehicle remote sensing image well The requirement of big data quantity processing, the defogging processing calculating speed of unmanned plane image further improve.Give full play to unmanned aerial vehicle remote sensing The characteristics of technology is quick, motor-driven, flexible, particularly in disaster remote sensing fields, lacking for traditional emergency method can be made up well Fall into.

2. improve cloud and mist image processing effect pockety.A kind of unmanned plane shooting image provided by the invention Optimize defogging processing method, by using the method for enhancing atmospheric transmissivity figure layer contrast, solve use well and help secretly Road priori defogging algorithm is handled cloud and mist image pockety, and image cloud and mist is distributed relatively thin part and occurred substantially " cross handle ", and cloud and mist is distributed thicker part and the problem of " undercuring " phenomenon that cloud and mist not yet removes completely then occurs, keeps away Exempt to influence picture quality due to " cross and handle " and " undercuring ", very high meeting is higher to picture quality in remote sensing application It is required that so that unmanned aerial vehicle remote sensing image provides technical support to important decision.It is highly relatively low to give full play to unmanned plane during flying, can be with The advantages of sensitive detection ground object or low latitude climate change, and unmanned plane shooting image scene information can be overcome to answer Difficulty miscellaneous, that the thickness degree of cloud and mist can also change with the change of scene depth, unmanned plane shooting image are distributed in cloud and mist Treatment effect when uneven is still preferable.

3. improve the defog effect of white portion similar to atmosphere light in picture.A kind of unmanned plane provided by the invention The optimization defogging processing method of shooting image, by two kinds of image enchancing methods of auto contrast and auto color, solve dark Channel prior defogging processing method can cause the brightness of image to decline, the problem of causing output image brightness to be less than input picture. Output image brightness is higher, when the radiation of light source intensity of some atural objects and atmospheric radiation intensity are close in scene, suchly Object area can also obtain satisfied defogging result.Status captured by unmanned aerial vehicle remote sensing image is complicated, even in the north Winter, which is likely to occur, is likely to occur substantial amounts of green house of vegetables and roof or water white in city in sheet of snowfield, farmland Mud road, it is also more satisfied using the dark channel prior defogging algorithm process effect of optimization.

4. reducing parameter setting, unmanned plane image defogging batch processing is realized.A kind of unmanned plane provided by the invention is clapped The optimization defogging processing method of image is taken the photograph, using foggy image adaptive defogging parameter in itself, reduces relevant parameter setting, solution Dark channel prior of having determined method carries out the more bottleneck problem of processing parameter setting to unmanned plane image.Due to the figure of unmanned plane shooting Picture map sheet is smaller, and amount of images is more, if being entered using dark channel prior defogging algorithm to the artificial setting quantity of parameters of each image Row defogging processing, cost of labor will be greatly increased and have a strong impact on operating efficiency, it is achieved that using foggy image in itself from It is particularly significant to adapt to defogging parameter, treatment effeciency can be greatly improved, realize unmanned plane image defogging batch processing.

Brief description of the drawings

Fig. 1 is a kind of flow frame diagram of the optimization defogging processing method of unmanned plane shooting image of the present invention.

Fig. 2 is to update core histogram in the dark filtering that the Algorithms T-cbmplexity of the present invention is O (1).

Fig. 3 is the data transfer flow schematic diagram of the present invention.

Embodiment

Below in conjunction with the accompanying drawings, to a kind of technology of the optimization defogging processing method of unmanned plane shooting image provided by the invention Scheme is further described, and those skilled in the art is better understood from the present invention and can be practiced.

Referring to Fig. 1 to Fig. 3, the optimization defogging processing method of a kind of unmanned plane shooting image provided by the invention, to help secretly Based on road priori defogging processing method, with reference to the quality of data after unmanned plane feature of image optimization calculating speed and defogging, obtain It is adapted to the defogging method of unmanned plane shooting image；

Dark channel prior defogging processing method obtains priori according to dark principle, utilizes foggy image degeneration physics The processing of model realization image defogging, is divided into the following steps according to foggy image degeneration physical model by image defogging processing procedure： Estimate atmosphere illumination intensity A, estimation atmospheric transmissivity, defogging processing；Mean filter side is inserted when image cloud and mist is uniform simultaneously Method, guiding filtering method is being inserted when image cloud and mist is uneven；

Enter with reference to the optimization defogging algorithm of unmanned plane feature of image in terms of processing speed optimization and treatment effect optimization two Hand, processing speed optimization include following three aspects：First, the dark that Algorithms T-cbmplexity is O (1) filters, second, using Image down sampling reduces data volume, third, being refined atmospheric transmissivity figure layer using medium filtering；Treatment effect optimization includes following Two aspects：When enhancing atmospheric transmissivity figure layer contrast, second, image enhancement processing.

First, dark channel prior defogging processing method

Dark channel prior defogging processing method is the statistical theory drawn on the basis of the clear picture of big discharge observation：Big Majority does not include in the clear picture of sky, the gray scale of each pixel at least one passage on r, tri- Color Channels of g, b Value is very low or even levels off to 0, and in other words, the minimum value of radiation intensity should be very low in the regional area of some very little. Clearly in the region in picture in addition to sky portion, gray value all very littles of nearly all dark even level off to 0, Such observation statistical conclusions are dark channel prior.

Due to the influence that atmosphere light is shone, foggy image compare that brightness is higher with clearly fog free images and transmissivity t compared with Low, so often the dark channel value of foggy image is higher, dark channel image gray value is also higher.Dark channel image is generally described The thickness degree of cloud and mist, it can further be solved by dark figure layer and obtain atmospheric transmissivity, but due to dark figure layer It is only preliminary to present cloud and mist situation in air, so the atmospheric transmissivity solved by it is not met by defogging processing Required precision to it, it is necessary to carry out further refined processing.

2nd, image degradation model

In computer vision and field of Computer Graphics, following formula is by widely as the degeneration for describing foggy image Model：

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

In above-mentioned model, I (x) represents the radiation signal that sensor receives, namely the foggy image of input；J (x) tables Show that scene radiates, namely the picture rich in detail after processing；A represents global atmospheric radiation；T (x) represents transmissivity, defogging processing Target is exactly that the foggy image received by sensor recovers fogless picture rich in detail.

Section 1 J (x) I (x) are direct attenuation term in above formula, and Section 2 A (1-t (x)) shines for atmosphere light, direct attenuation term Describe scene radiation and scene radiates the loss of signal amount in communication process, atmosphere light is big received by image according to referring to Gas scattered light intensity is also an important factor for causing color of image change.

3rd, atmosphere illumination intensity A is estimated

Prior art estimates the method for atmosphere illumination intensity all by the gray value of all pixels in image by single image Maximum as atmosphere illumination intensity, but in real image, it is usually possible to be figure for the maximum pixel of gray value The white object of other in piece, such as the automobile or house of white.As dark channel image generally depicts the thickness of cloud and mist Degree, so utilizing dark channel image estimation atmosphere illumination intensity so that the atmosphere illumination intensity of estimation is closer to actual value.

Because dark channel image depicts the thickness degree of cloud and mist, gray value is higher in dark channel image represents region cloud and mist It is thicker, the gray value of pixel of this subregion is located in the foggy image of input closer to atmosphere illumination intensity, is estimated Atmosphere illumination intensity is calculated firstly the need of finding gray value highest in dark channel image and account for the pixel of image total pixel number amount 0.1% Point, its position is recorded, 10 pictures before gray value highest need to be found out comprising 10000 pixels even in dark channel image Element simultaneously records its corresponding coordinated indexing in the picture, found according to the coordinated indexing of record in the foggy image of input with Its corresponding pixel, and these pixel gray level average values found in foggy image are calculated as atmosphere illumination intensity.

4th, atmospheric transmissivity is estimated

Atmospheric transmissivity is considered as invariant in the filter window centered on pixel x when estimating atmospheric transmissivity, made The atmospheric transmissivity in the filter window centered on pixel x is represented with t (x), calculates the minimum in each Color Channel respectively Value, calculates the minimum value of gray scale in the filter window centered on pixel x, then air projection ratio can directly be calculated.

Now because dark channel image can only substantially show cloud and mist thickness situation, so being calculated according to dark channel image Atmospheric transmissivity be not met by the requirement of defogging processing accuracy, if by force using its carry out defogging processing can out " halation " it is existing As the reason for " halation " phenomenon occurs block operation for being used when dark is being asked for of key factor, asking for dark figure layer When, the dark channel value of some pixel is all 3 passages of all pixels in the filter window of the n*n centered on the pixel Minimum gradation value, if there is the gray value of an a certain passage of pixel very low in image, then when calculating dark figure layer, All can be relatively low with the dark channel value of pixel all in the range of the centrical n*n of pixel filter window so that color It is dimmed, cause to expand phenomenon, next how to refine to transmissivity figure layer to discuss.

5th, defogging is handled

The defogging method of dark channel prior has become the main flow of image defogging at present, and this method is based on dark channel prior On basis, the thickness depth of cloud and mist and the radiation reduction amount of each pixel can be quantified based on this priori conditions, from And reduce the picture rich in detail of high quality.

According to foggy image degradation model, the purpose of image defogging processing is to recover defeated by the foggy image I inputted Go out clearly fog free images J.It can be obtained by foggy image degradation model formula：

By the atmospheric transmissivity t (x) in the filter window centered on pixel x, foggy image I (x), atmosphere illumination intensity A, you can draw fog free images J (x).

6th, the dark that Algorithms T-cbmplexity is O (1) filters

Dark channel prior refers to that the minimum value of radiation intensity in the regional area of some very little should be very low, then dark The minimum value of radiation intensity in filter window region is found in filtering using the form of filtering.It is black to regard dark filtering as two dimension Extension of the mini-value filtering of white image to three dimensions coloured image, three dimensions refer to r, g, b color spaces, dark Filtering will look for r in spectral window mouth region, g, the minimum pixel of radiation intensity and make the radiation intensity of the pixel in b color spaces For dark channel value.

Mini-value filtering is a kind of typical image manipulation, but higher algorithm complex has had a strong impact on minimum value filter The application of ripple device, because the algorithm has the characteristics of non-linear and inseparability common optimisation technique to it and does not apply to, All it is generally using image pixel as a list, wherein minimum value, main minimum is then ranked up to it The algorithm complex of value filtering changes with the difference of sort method, if the algorithm using bucket sort so mini-value filtering Complexity is O (n^2), and wherein n is filter window size.

A kind of unrelated maximin filtering algorithm of window size proposed by the present invention, it is cumulative due to two histograms It is the operation of O (1), so as to per an increased pixel, histogram need to only add up a fixed number, so as to obtain O (1) complexity, specifically, Algorithms T-cbmplexity are that O (1) mini-value filtering algorithm is made up of the following steps：

1. initialize core histogram.If image has n row, then we just need to safeguard n histogram, each Nogata Figure in processing procedure, will be updated, these histograms include for each row in image to the histogram of each row The information of the 2r+1 pixel centered on some pixel, original state are 2r+1 picture centered on each first pixel of row The histogram information of element, in processing, these histograms are updated simultaneously with the movement of sliding block.

2. update core histogram.More new histogram, it is as sliding block moves, leftmost side picture is subtracted to the histogram of each row The histogram information of element, while increase the histogram information of the pixel of the rightmost side, this step is also O (1) operation, such as Fig. 2 institutes Show.

3. obtain minimum value.Minimum value is taken according to the core histogram of renewal.

Pass through above-mentioned algorithm calculating process, it can be seen that all single pixel operations are all the i.e. calculations unrelated with window size Method space complexity is O (1), and the mini-value filtering of above-mentioned gray level image is extended and is applied to coloured image, can First to carry out mini-value filtering to r, the figure layer of tri- wave bands of g, b respectively, compare take r afterwards, tri- wave band figure layers of g, b were entered most Each pixel after small value filtering, take minimum value in three pixel values as dark channel value, realize with filter window without The dark that pass Algorithms T-cbmplexity is O (1) filters.

7th, data volume is reduced using image down sampling

In computer graphics and computer vision field, the process being adjusted to image size is referred to as image scaling, Need to do a balance in the smoothness and definition and treatment effeciency of this result.When image down, image is smooth Degree and definition will increase and data volume will be compressed, on the contrary, when the size increase of image, image data amount will Increase and smoothness and clear figure will decline, sampling principle is as shown in Figure 3.

The process of downscaled images is referred to as image down sampling, and the process of enlarged drawing is then referred to as up-sampling.In actual applications The main purpose of image down sampling has two：First, in real world images, downscaled images size is allowed to meet viewing area big It is small；Second, the thumbnail of correspondence image is produced, reduces image data amount.The main purpose of picture up-sampling is mainly being shown Image is amplified during image, so as to be shown on the higher display device of resolution ratio.

In dark channel prior algorithm, dark figure layer is used for rudenss quantization cloud and mist thickness, and is entered using dark channel image One step calculates rough atmospheric transmissivity, and then atmospheric transmissivity figure layer is become more meticulous using the methods of guiding filtering.Secretly Channel image serves to form a connecting link in whole defogging algorithm It is very high, another aspect dark channel image in calculating process very easily by noise effect in image, so we are natural The mode expected using image down sampling reduce input picture data volume generation input picture thumbnail, utilize input figure As the thumbnail of thumbnail calculating dark channel image, finally the thumbnail of dark channel image is entered using the method used on image Row amplification, further to be calculated.It so can not only reduce noise in input picture to calculating the shadow of dark channel image Ring, while decrease data amount of calculation, reduce and calculate the time.

The processing experiment of image defogging is carried out using the cross-platform computer vision figures of increasing income of OpenCV in the present embodiment, is utilized OpenCV realizes to be operated to the up-sampling and down-sampling of input picture, can simply can be respectively to image using two functions Up-sampled and down-sampling.We carry out down-sampling using gaussian pyramid to image herein, use Laplce's gold word The image of lower floor's low resolution is carried out up-sampling reconstruction by tower.

8th, refined atmospheric transmissivity figure layer using medium filtering

Transmissivity mentioned above using rough estimate, which directly carries out defogging processing, will occur " halation " phenomenon, and its is main Reason is：When asking for dark figure layer, the dark channel value of some pixel is all the filter of the n*n centered on the pixel The minimum gradation value of 3 passages of all pixels in ripple window, if having in image, the gray value of an a certain passage of pixel is very low, So when calculating dark figure layer, with the dark of pixel all in the range of the centrical n*n of pixel filter window Channel value all can be relatively low so that colour-darkening, causes to expand phenomenon.

In order to solve " halation " phenomenon, a suitable wave filter is chosen, it can realize and is smoothed to image While protect image edge information, be the key solved the problems, such as, although median filter to protection effect for image edges not It is prominent, but it is the simplest wave filter with edge-protected effect, realizes that the intermediate value that Algorithms T-cbmplexity is O (1) is filtered Ripple device, implemented simply compared to median filter for guiding filtering, Algorithms T-cbmplexity is relatively low, can be largely Reduce the calculating time of defogging algorithm.

9th, atmospheric transmissivity figure layer contrast is strengthened

When asking for atmospheric transmissivity, for the picture for having mist, atmospheric transmissivity and image depth inversely, That is the bigger atmospheric transmissivity of the scene depth of field is smaller.For the dark figure layer of foggy image, the thicker local image of fog Pixel radiation intensity is bigger in the region, and vice versa.When asking for atmospheric transmissivity, the dark channel diagram of foggy image is first asked for As simultaneously being normalized to obtain image G (x), image G (x) is taken and inverse obtains atmospheric transmissivity figure layer：

T (x)=1-G (x)

A parameter ∈ is usually added into actual applications to be used to adjust atmospheric transmissivity：

T (x)=1- ∈ G (x)

Find that ∈ span is preferred between 0.9~1.2 after many experiments.

The higher pixel region cloud and mist of gray value is thicker in dark channel image, and atmospheric transmissivity is relatively low, in defogging These pixel region radiation compensation amounts are larger in algorithm, so atmospheric transmissivity figure layer is well to cloud and mist thickness and image Required radiation compensation is tentatively quantified, but because image gray levels are only 256 grades, so often being counted to foggy image Obtained atmospheric transmissivity figure layer pixel grey scale is always concentrated between 150~255, and gradation of image information is heavily compressed, special Be not in the case of image cloud and mist skewness, because atmospheric transmissivity figure layer gray value is concentrated in a certain region, that The scope that a very little is compressed to the radiation reduction amount of entire image is meant that, for the relatively thin region of cloud and mist and cloud The radiation reduction amount difference in the thicker region of mist is little, and radiation reduction amount is excessive for region so relatively thin to cloud and mist, will Cause " cross and handle " phenomenon, cause cloud and mist to go completely conversely, for the thicker area radiation of cloud and mist also commercial weight is smaller Remove, the image processing effect that mainly cloud and mist can be caused uneven substantially reduces, and has a strong impact on picture quality after processing.

In order to improve cloud and mist it is uneven in the case of, picture quality after dark defogging algorithm process, to atmospheric transmissivity figure Layer carries out contrast stretching.Contrast stretching can improve the dynamic range of image.It is total can so image pixel gray level to be increased The dynamic range of the dark channel image concentrated between 150~255, the thicker region in the relatively thin region of cloud and mist and cloud and mist is distinguished Treat so that the relatively thin area radiation of cloud and mist also commercial weight reduces, there is provided the radiation reduction amount of cloud and mist thicker region, so as at raising Picture quality after reason.Method the present embodiment that gray scale is drawn high uses linear stretch method.

Tenth, image enhancement processing

Dark channel prior defogging processing method can cause the brightness of image to decline, and cause output image brightness to be less than input figure Picture, so needing to carry out grey level enhancement processing to output image.The present embodiment uses two kinds of figures of auto contrast and auto color Image intensifying method.

1. auto contrast

Because auto contrast does not adjust passage individually, so will not increase or eliminate colour cast problem, auto contrast's life Most bright in image and most dark pixel is mapped to white and black by order so that the higher region of gray scale is brighter, the relatively low area of gray scale Domain is darker.Auto contrast's order realizes that step is as follows：

The first step, set for determining parameter a HighCut and LowCut for passage upper lower limit value；

Second step, to r, the image of tri- passages of g, b, distribution statisticses grey level histogram；

3rd step, respectively to the image setting parameter of each passage, so that it is determined that upper lower limit value, wherein N and M are distributed as image It is wide and high；

4th step, compare the maximum and minimum value of three passage upper lower limit values, with the minimum value of lower limit and the maximum of the upper limit It is newly upper and lower bound that value, which is used as, and calculates and insinuate table；

5th step, root tool insinuate table, r, the image of tri- passages of g, b are insinuated respectively.

2. auto color

Auto color operation be that the contrast of image is adjusted according to the actual grey of image, in practical operation I Shadow and highlight value in clip image, then the most bright and most dark pixel of the remainder in image is mapped to pure white or pure Black, intermediate pixel redistributes distribution in proportion.

Auto color realizes that first three step is identical with auto contrast in step, but is three in auto contrast's calculating process Passage builds different tables of insinuating respectively, then by three it is different insinuate table based on three wave bands are insinuated.

11, experiment is compared and conclusion

We use the objective image evaluation method without reference in an experiment, and this kind of method has independent of reference picture The characteristics of, objective evaluation can be carried out to image in the case of without reference to image.It is strong using image information entropy, image border Degree and three evaluation indexes of image variance.Allocation of computer is as follows used in this experiment：

Processor：Core^TMi7-5500U@2.4GHz；

Internal memory (RAM) is installed：8.00GB；

System type：The bit manipulation systems of Win 8.1 64, the processor based on x64；

It is as follows using data message in this experiment：

Spot for photography：Image is unmanned plane in aerophotograph captured by Tianjin；

Map sheet size：5780*5400pix；

Picture format：.jpg form；

Color Channel：Including r, tri- passages of g, b；

1st, dark channel image computing module speed-optimization

First, dark defogging algorithm improves the definition of image really, fog in foggy image is removed, dark Priori defogging algorithm effect it is further preferred that；Secondly, when calculating dark figure layer, reduce data volume using image down sampling and use algorithm Time complexity filters two kinds of prioritization schemes by O (1) dark and dark channel image is calculated, with using conventional method meter Obtained dark channel image difference is little, and the dark channel image that operating speed prioritization scheme is calculated is to defogging treatment effect Do not have an impact；Finally, for dark channel image calculating speed, data volume is reduced and using calculation using image down sampling The dark that method time complexity is O (0) filters the calculating speed that dark channel image has been significantly increased in two kinds of prioritization schemes very much Degree, 180 times are improved compared to traditional dark computational methods speed.

2nd, atmospheric transmissivity figure layer is refined module speed-optimization

In the case of cloud and mist is uniform, although replacing guiding filtering to carry out essence to atmospheric transmissivity figure layer using medium filtering Defog effect has been influenceed after change, but can still reduce clearly image, has but obtained carrying greatly very much in terms of calculating speed Height, speed improves nearly 9 times than speed before optimization after optimization, can be in use so on the premise of image use is not influenceed Value filtering is refined instead of guiding filtering to atmospheric transmissivity figure layer.

3rd, dark channel prior defogging algorithm quality optimization is tested

In the case where cloud and mist is uneven, contrast stretching is carried out to atmospheric transmissivity figure layer and can be very good to improve defogging Effect, obtain the image become apparent from.But due to the step of adding contrast stretching, under will necessarily causing calculating speed Drop, but the time is calculated still within the scope of rational.

4th, output image enhancing processing experiment

First, can be darker than original foggy image using the image after the direct defogging of method of dark channel prior；Secondly, regarding In feel, the visual effect of mist elimination image can be improved using auto contrast and auto color Enhancement Method；Further, using automatic Contrast and auto color can effectively improve picture quality, can particularly strengthen the marginal information and contrast of image. In summary, enhancing processing is carried out to output image using auto contrast and auto color, can effectively improves dark Priori defogging algorithm output image quality.

The present invention is based on dark channel prior defogging Processing Algorithm, for unmanned plane image data amount is big, more than map sheet Feature, propose algorithm speed prioritization scheme.For reducing two aspect problems of data volume and optimized algorithm time complexity, propose Data volume method and Algorithms T-cbmplexity are reduced using image down sampling to filter for O (1) dark；Concurrently referring now to figure The uniform situation of cloud and mist as in, can be refined using medium filtering instead of guiding filtering to atmospheric transmissivity figure layer, so as to Reduce and calculate the time.By speed-optimization, algorithm has well adapted to the characteristics of unmanned plane image data amount is big.Due to unmanned plane Aerial Images are mainly used in agricultural, geology, hydrological environment monitoring and military surveillance etc., so for image quality requirements Higher, the present invention proposes two kinds of prioritization schemes for this problem：It is right for the uneven image of cloud and mist during defogging Transmissivity figure layer carries out contrast stretching；Complete to carry out enhancing processing to image after defogging processing.By quality optimization, nobody Machine image defogging Processing Algorithm output image quality has obtained effective raising.

The basic principles, principal features and advantages of the present invention have been shown and described above.While there has been shown and described that Embodiments of the invention, for the ordinary skill in the art, it is possible to understand that do not departing from the principle and essence of the present invention A variety of change, modification, replacement and modification can be carried out to these embodiments, the scope of the present invention is by appended right in the case of god It is required that and its equivalent restriction.

Claims (10)

1. the optimization defogging processing method of a kind of unmanned plane shooting image, it is characterised in that with dark channel prior defogging processing side Based on method, with reference to the quality of data after unmanned plane feature of image optimization calculating speed and defogging, obtain being adapted to unmanned plane shooting figure The defogging method of picture；

The dark channel prior defogging processing method obtains priori according to dark principle, utilizes foggy image degeneration physics The processing of model realization image defogging, is divided into the following steps according to foggy image degeneration physical model by image defogging processing procedure： Estimate atmosphere illumination intensity A, estimation atmospheric transmissivity, defogging processing；

Start with reference to the optimization defogging algorithm of unmanned plane feature of image in terms of processing speed optimization and treatment effect optimization two, locate Managing speed-optimization includes following three aspects：First, the dark that Algorithms T-cbmplexity is O (1) filters, second, using image Down-sampling reduces data volume, third, being refined atmospheric transmissivity figure layer using medium filtering；Treatment effect optimization includes following two Aspect：When enhancing atmospheric transmissivity figure layer contrast, second, image enhancement processing.

2. the optimization defogging processing method of a kind of unmanned plane shooting image according to claim 1, it is characterised in that help secretly Road principle is the statistical theory drawn on the basis of the clear picture of big discharge observation：In most of clear pictures for not including sky In, the gray value of each pixel at least one passage on r, tri- Color Channels of g, b is very low or even levels off to 0, at some The minimum value of radiation intensity is very low in the regional area of very little, foggy image compared with clearly fog free images brightness it is higher and Transmissivity t is relatively low, so the dark channel value of foggy image is higher, dark channel image gray value is also higher, and dark channel image is described The thickness degree of cloud and mist, can solve to obtain atmospheric transmissivity by dark figure layer.

A kind of 3. optimization defogging processing method of unmanned plane shooting image according to claim 1, it is characterised in that estimation Atmosphere illumination intensity A finds gray value highest in dark channel image and accounts for the pixel of image total pixel number amount 0.1% first, note Its position is recorded, and records its corresponding coordinated indexing in the picture, mist is had in input according to the coordinated indexing of record Corresponding pixel is found in image, and calculates these pixel gray level average values found in foggy image as big Gas intensity of illumination.

A kind of 4. optimization defogging processing method of unmanned plane shooting image according to claim 1, it is characterised in that estimation Atmospheric transmissivity is considered as invariant in the filter window centered on pixel x during atmospheric transmissivity, using t (x) represent with The atmospheric transmissivity in filter window centered on pixel x, calculates the minimum value in each Color Channel, calculates with picture respectively The minimum value of gray scale in filter window centered on plain x, then air projection ratio can directly be calculated.

A kind of 5. optimization defogging processing method of unmanned plane shooting image according to claim 1, it is characterised in that according to Foggy image degradation model, the processing of image defogging recovers output clearly fog free images J by inputting foggy image I, under Formula

<mrow> <mi>J</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mrow> <mi>A</mi> <mrow> <mo>(</mo> <mn>1</mn> <mo>-</mo> <mi>t</mi> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>)</mo> <mo>-</mo> <mi>I</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> </mrow> <mrow> <mi>t</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> </mrow> </mfrac> </mrow>

By the atmospheric transmissivity t (x) in the filter window centered on pixel x, foggy image I (x), atmosphere illumination intensity A, i.e., Fog free images J (x) can be drawn.

A kind of 6. optimization defogging processing method of unmanned plane shooting image according to claim 1, it is characterised in that algorithm Time complexity be O (1) in O (1) dark filtering method be using the algorithm complex of the mini-value filtering of bucket sort, Wherein 1 is filter window size, because two the cumulative of histogram are O (1) operations, so as to per increased one Pixel, histogram need to only add up a fixed number, so as to obtain the complexity of O (1), specifically, Algorithms T-cbmplexity For O (1) mini-value filtering algorithm by forming in several steps：

The first step, initialize core histogram；If image there is n to arrange, it is necessary to safeguard n histogram, each histogram is for figure Each row as in, in processing procedure, will be updated, these histograms include with some picture to the histogram of each row The information of 2r+1 pixel centered on element, r are filter window radius, and original state is centered on each first pixel of row The histogram information of 2r+1 pixel, in processing, these histograms are updated simultaneously with the movement of sliding block；

Second step, update core histogram；More new histogram is as sliding block moves, and leftmost side picture is subtracted to the histogram of each row The histogram information of element, while increase the histogram information of the pixel of the rightmost side, this step is also O (1) operation；

3rd step, obtain minimum value；Minimum value is taken according to the core histogram of renewal；

All single pixel operations are all unrelated with window size, and the mini-value filtering of above-mentioned gray level image is extended and fitted For coloured image, first are carried out by mini-value filtering, compares take r afterwards for r, the figure layer of tri- wave bands of g, b respectively, tri- ripples of g, b Each pixel that section figure layer was entered after mini-value filtering, the minimum value in three pixel values is taken to be realized as dark channel value The dark that Algorithms T-cbmplexity unrelated with filter window is O (1) filters.

7. the optimization defogging processing method of a kind of unmanned plane shooting image according to claim 1, it is characterised in that use It is to reduce the data volume generation input picture of input picture using the mode of image down sampling that image down sampling, which reduces data volume, Thumbnail, the thumbnail of dark channel image is calculated using input picture thumbnail, finally will be dark using the method used on image The thumbnail of channel image is amplified, further to be calculated.

8. the optimization defogging processing method of a kind of unmanned plane shooting image according to claim 1, it is characterised in that use Medium filtering atmospheric transmissivity figure layer of refining solves halation phenomenon, and median filter protects while being smoothed to image Protect image edge information.

A kind of 9. optimization defogging processing method of unmanned plane shooting image according to claim 1, it is characterised in that enhancing Atmospheric transmissivity figure layer contrast is that the dark channel image of foggy image is first asked for when asking for atmospheric transmissivity and is carried out Normalized obtains image G (x), and image G (x) is taken against atmospheric transmissivity figure layer is obtained, is used to adjust wherein adding parameter ∈ Atmospheric transmissivity is saved, ∈ span is [0.9,1.2]：

T (x)=1- ∈ G (x)

To the image after dark defogging algorithm process, atmospheric transmissivity figure layer contrast stretching is carried out.

A kind of 10. optimization defogging processing method of unmanned plane shooting image according to claim 1, it is characterised in that figure Image intensifying processing method includes auto contrast and auto color；

Most bright in image and most dark pixel is mapped to white and black by auto contrast, and the higher region of gray scale is brighter, gray scale Relatively low region is darker, and auto contrast realizes that step is as follows：

1st step, set for determining parameter a HighCut and LowCut for passage upper lower limit value；

2nd step, to r, the image of tri- passages of g, b, distribution statisticses grey level histogram；

3rd step, respectively to the image setting parameter of each passage, so that it is determined that upper lower limit value, wherein N and M are distributed as the width of image And height；

4th step, compare the maximum and minimum value of three passage upper lower limit values, using the maximum of the minimum value of lower limit and the upper limit as It is newly upper and lower bound, and calculates and insinuate table；

5th step, root tool insinuate table, r, the image of tri- passages of g, b are insinuated respectively；

Auto color is that the contrast of image is adjusted according to the actual grey of image, the shadow and highlight in clip image Value, then the most bright and most dark pixel of the remainder in image is mapped to pure white or black, intermediate pixel divides again in proportion With distribution.