CN106469440B - Dark defogging parallel optimization method based on OpenCL - Google Patents
Dark defogging parallel optimization method based on OpenCL Download PDFInfo
- Publication number
- CN106469440B CN106469440B CN201610808183.9A CN201610808183A CN106469440B CN 106469440 B CN106469440 B CN 106469440B CN 201610808183 A CN201610808183 A CN 201610808183A CN 106469440 B CN106469440 B CN 106469440B
- Authority
- CN
- China
- Prior art keywords
- image
- gray
- transmittance
- dark
- defogging
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 24
- 238000005457 optimization Methods 0.000 title claims abstract description 12
- 238000002834 transmittance Methods 0.000 claims abstract description 44
- 238000012545 processing Methods 0.000 claims abstract description 24
- 238000010586 diagram Methods 0.000 claims abstract description 15
- 239000003595 mist Substances 0.000 claims abstract description 14
- 239000011159 matrix material Substances 0.000 claims description 33
- 238000001914 filtration Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 3
- 230000015572 biosynthetic process Effects 0.000 claims description 2
- AGVAZMGAQJOSFJ-WZHZPDAFSA-M cobalt(2+);[(2r,3s,4r,5s)-5-(5,6-dimethylbenzimidazol-1-yl)-4-hydroxy-2-(hydroxymethyl)oxolan-3-yl] [(2r)-1-[3-[(1r,2r,3r,4z,7s,9z,12s,13s,14z,17s,18s,19r)-2,13,18-tris(2-amino-2-oxoethyl)-7,12,17-tris(3-amino-3-oxopropyl)-3,5,8,8,13,15,18,19-octamethyl-2 Chemical compound [Co+2].N#[C-].[N-]([C@@H]1[C@H](CC(N)=O)[C@@]2(C)CCC(=O)NC[C@@H](C)OP(O)(=O)O[C@H]3[C@H]([C@H](O[C@@H]3CO)N3C4=CC(C)=C(C)C=C4N=C3)O)\C2=C(C)/C([C@H](C\2(C)C)CCC(N)=O)=N/C/2=C\C([C@H]([C@@]/2(CC(N)=O)C)CCC(N)=O)=N\C\2=C(C)/C2=N[C@]1(C)[C@@](C)(CC(N)=O)[C@@H]2CCC(N)=O AGVAZMGAQJOSFJ-WZHZPDAFSA-M 0.000 claims description 2
- 238000003860 storage Methods 0.000 claims description 2
- 238000012544 monitoring process Methods 0.000 abstract description 5
- 230000007547 defect Effects 0.000 abstract description 2
- 230000003287 optical effect Effects 0.000 abstract 1
- 238000002474 experimental method Methods 0.000 description 7
- HPTJABJPZMULFH-UHFFFAOYSA-N 12-[(Cyclohexylcarbamoyl)amino]dodecanoic acid Chemical compound OC(=O)CCCCCCCCCCCNC(=O)NC1CCCCC1 HPTJABJPZMULFH-UHFFFAOYSA-N 0.000 description 6
- 238000012360 testing method Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 238000009738 saturating Methods 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000007123 defense Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012806 monitoring device Methods 0.000 description 1
- 238000011056 performance test Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/73—Deblurring; Sharpening
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T1/00—General purpose image data processing
- G06T1/20—Processor architectures; Processor configuration, e.g. pipelining
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
Abstract
The dark defogging parallel optimization method based on OpenCL that the invention discloses a kind of it is limited mainly to solve the applicable platform of the prior art, it is difficult to the defect that high-definition picture is handled in real time, implementation step are as follows: the initialization of 1. equipment；2. will have a frame image copy of mist video into equipment；3. calculating the grayscale image and triple channel minimum value figure of foggy image, and calculate dark channel diagram；4. calculating atmosphere light according to dark channel diagram；5. estimating transmittance figure according to dark channel diagram and atmosphere optical oomputing；Fine transmittance figure is obtained 6. pair estimating transmittance figure and carrying out Steerable filter；7. according to atmosphere light, foggy image and fine transmittance figure carry out defogging processing, and by the image copy after defogging to host memory.The present invention is suitable for a variety of parallel computing platforms, improves the processing speed of dark defogging, can be used for for subsequent video monitoring, and image recognition provides clearly image source.
Description
Technical field
The invention belongs to technical field of image processing, are further a kind of defogging parallel optimization methods, can be used for rear
Continuous real-time video monitoring, image recognition provide clearly image source.
Background technique
Under the conditions of haze weather, outdoor images acquisition equipment is affected by it, and acquired image will appear fuzzy, contrast
Decline, the degradation phenomenas such as hue shift have not only seriously affected the normal work of vision detection system, also give later period intelligent recognition
Huge challenge is brought Deng application.
The dark channel prior defogging method that He Kaiming et al. is proposed, has obtained academia in terms of single-frame images defog effect
Approval since to calculate the time longer for this method, pass through simple CPU serial arithmetic, it is difficult to full however in terms of processing speed
Requirement of the sufficient video monitoring scene to real-time.
Be conducive to segmentation based on image data, can a large amount of parallel feature, many solutions of the Heterogeneous Computing based on CPU+GPU
Certainly scheme is come into being." the dark primary priori defogging algorithm based on CUDA is simultaneously in paper by Xue Yungang of the National University of Defense technology et al.
Row realize and optimization " in propose a kind of Parallel Implementation method based on CUDA Computational frame, to 640 × 480 image have 25 frames/
The processing speed of second；Sichuan Chuandazhisheng Software Co., Ltd, a kind of patent application " figure based on CUDA of Sichuan University
As real-time defogging method " (publication number: CN103049890A, application number: CN201310017014.X, the applying date: in January, 2013
17 days) in disclose a kind of method of real-time defogging of the image based on CUDA, 32 frames/second can achieve to 600 × 400 image
Processing speed.These methods have the disadvantage that
Firstly, existing Heterogeneous Computing solution reaches the processing speed of the 25 frames/more than second to low resolution, with
In recent years monitoring device gradually to 1920 × 1080 high-resolution transition, this unappeasable HD video of processing speed
Real-time processing requirement；
Secondly, be only capable of on the GPU for operating in the production of NVIDIA company based on the program that CUDA Computational frame is write, it is removable
Plant property is lower.
Summary of the invention
It is an object of the invention to overcome the shortcomings of prior art, propose that a kind of dark defogging based on OpenCL is parallel
Optimization method, to improve the defogging speed of video and realize the compatibility of platform.
To achieve the goals above, the present invention includes:
(1) platform for supporting open computational language OpenCL is selected, which is initialized, and therefrom chooses and can transport
The equipment E of row kernel function, the equipment include CPU, GPU, intel MIC accelerator card；
(2) there is a frame image I of mist video from the global memory that host memory copies equipment E to RGB color；
(3) gray value of each location of pixels of RGB color foggy image I is calculated by way of parallel processing, generates gray scale
Scheme Igray, each location of pixels RGB triple channel minimum value is calculated by way of parallel processing, generates minimum value figure Imin；
(4) to minimum value figure IminMini-value filtering is carried out, dark channel diagram I is obtaineddark；
(5) according to dark channel diagram IdarkCalculate atmosphere light A:
(5a) calculates dark channel diagram IdarkHistogram h；
(5b) acquires the corresponding gray value lower bound l of maximum 0.1% pixel of gray value according to histogram h；
(5c) is threshold value to dark channel image I using ldarkIt carries out binaryzation and obtains bianry image I'；
The RGB triple channel figure I of (5d) to foggy image Ir、Ig、IbDot product bianry image I' respectively obtains RGB triple channel ash
Spend image Ir'、Ig'、Ib'；
(5e) calculates separately RGB triple channel gray level image Ir'、Ig'、Ib' histogram hr、hg、hb；
(5f) passes through RGB three channel histogram hr、hg、hb, calculate atmosphere light A={ Ar,Ag,Ab,
Wherein,C ∈ { r, g, b }, i are the gray value in 0~255 variation, Ar、Ag、AbIt is respectively red
Color, green, the air light value of blue triple channel；
(6) according to the RGB triple channel figure I of atmosphere light A and foggy image Ir、Ig、Ib, calculate and estimate transmittance figure t；
(7) according to the grayscale image I of foggy imagegrayWith estimate transmittance figure t, obtain fine transmittance figure t:
(7a) is gone to transmittance figure t is estimated to mean filter, and row mean filter result figure t is obtainedr；Again to row mean value
Filter result figure trIt is arranged to mean filter, obtains smooth transmissive rate figure mt；
The grayscale image I of (7b) to foggy imagegrayMean filter is carried out, smooth guide figure m is obtainedI；
(7c) is according to smooth transmissive rate figure mt, smooth guide figure mI, estimate the grayscale image of transmittance figure t and foggy image
Igray, obtain smooth scaled matrix ma, smoothed offset matrix mb；
(7d) is according to smooth scaled matrix ma, smoothed offset matrix mbWith the grayscale image I of foggy imagegray, obtain fine saturating
Penetrate rate figure
(8) according to RGB triple channel foggy image I, fine transmittance figureWith atmosphere light A, defogging result figure J is calculated；
(9) RGB triple channel defogging result figure J is read back host memory from equipment E；
The present invention compared with prior art, has the advantages that
First, it, can be with based on the program that OpenCL language is write due to present invention uses open computational language OpenCL
It is run in the MIC platform of INTEL, the CPU of AMD, NVIDIA, the GPU of AMD, INTEL, on the other hand, other solutions are adopted
The CUDA computational language used is only capable of running on NVDIA video card as the exclusive Computational frame of NVIDIA, therefore, the present invention program
With better platform compatibility.
Second, since present invention introduces histograms to solve atmosphere light, calculated with the former dark defogging for choosing single pixel
Method is compared, and gained air light value numerically has stability.
Third, since the present invention has used local memory in mean filter, the type memory has very high access speed
It spends, the use in the present invention to local memory, ensure that the read-write operation to global memory to merge access, improves reading and writes
Enter efficiency, saves the processing time.
4th, it is calculated since the present invention uses first piecemeal in mean filter, the method for remerging piecemeal processing result,
Degree of parallelism is improved, the processing time is saved.
5th, since the present invention uses advanced every trade to filtering in mean filter, then arranged the ranks to filtering
Isolated processing method avoids the redundant access to same pixel, improves processing speed.
Detailed description of the invention
Fig. 1 is implementation flow chart of the invention.
Fig. 2 is the frame image that the RGB color that the present invention uses has mist video.
Fig. 3 is defogging result figure of the invention.
Specific embodiment
Referring to the drawings, the specific embodiment of the invention and effect are described further:
Referring to Fig.1, the present invention uses OpenCL Computational frame, can be in any a equipment for supporting OpenCL framework
It realizes, its step are as follows:
Step 1, it initializes.
The platform for supporting open computational language OpenCL is selected, which is initialized, and therefrom chooses and can run
The equipment E of kernel function, the equipment include CPU, GPU, intel MIC accelerator card.
Step 2, there is a frame image I of mist video from the global memory that host memory copies equipment E to RGB color.
Since video is considered as the image sequence of Time Continuous, the present invention is by having mist video successive frame to RGB color
It carries out image defogging and realizes video defogging, the global memory that will there is mist video sequential frame image to copy equipment from host memory to
In, to carry out image defogging.
Step 3, grayscale image and triple channel minimum value figure are calculated.
With single thread process single pixel point, different threads handle the mode of different coordinate position pixels, calculate RGB
The gray value of colored each coordinate position of foggy image I forms grayscale image Igray；
Wherein, grayscale image IgrayPixel value at coordinate x:
Igray(x)=0.299 × r+0.587 × g+0.114 × b, r, g, b be respectively red of the foggy image I at x,
Green, blue pixel value；
With single thread process single pixel point, different threads handle the mode of different coordinate position pixels, calculate each
Location of pixels RGB triple channel minimum value forms minimum value figure Imin；
Wherein, minimum value figure IminPixel value at coordinate x: Imin(x)=min (r, g, b).
Step 4, to minimum value figure IminMini-value filtering is carried out, dark channel diagram I is obtaineddark。
Calculate dark channel diagram IdarkPixel value I at coordinate xdark(x):
Form dark channel diagram Idark, wherein Ω (x) is the square neighborhood centered on coordinate x.
Step 5, atmosphere light is calculated.
(5a) calculates dark channel diagram IdarkHistogram h；
(5b) acquires the corresponding gray value lower bound l of maximum 0.1% pixel of gray value according to histogram h:
Calculate accumulative and value of the array s at coordinate o:Meet following formula by accumulative and array s lookup
Gray value lower bound l:
S [l] >=0.001 × n
s[l+1]<0.001×n
Wherein, s [l], s [l+1] are respectively value of the s at coordinate l, l+1, and n is dark channel diagram IdarkPixel sum；
(5c) is threshold value to dark channel image I using ldarkBinaryzation is carried out, bianry image I' is obtained；
The RGB triple channel figure I of (5d) to foggy image Ir、Ig、IbDot product bianry image I' respectively obtains RGB triple channel ash
Spend image Ir'、Ig'、Ib'；
(5e) calculates separately RGB triple channel gray level image Ir'、Ig'、Ib' histogram hr、hg、hb；
(5f) passes through RGB three channel histogram hr、hg、hb, calculate atmosphere light A={ Ar,Ag,Ab,
Wherein,C ∈ { r, g, b }, i are the gray value in 0~255 variation, Ar、Ag、AbIt is respectively red
Color, green, the air light value of blue triple channel.
Step 6, it calculates and estimates transmittance figure.
(6a) is according to red, green, the air light value A of blue triple channelr、Ag、Ab, calculate and estimate transmittance figure t in coordinate
Pixel value t (z) at z:
Wherein Ir(y)、Ig(y)、IbIt (y) is respectively RGB triple channel figure Ir、Ig、IbGray value at coordinate y, Ω (z) are
Square neighborhood centered on coordinate z；
Pixel value t (z) at each coordinate is combined by (6b), and transmittance figure t is estimated in formation.
Step 7, to transmittance figure progress Steerable filter is estimated, fine transmittance figure is obtained.
(7a) is gone to transmittance figure t is estimated to mean filter, and row mean filter result figure t is obtainedr；Again to row mean value
Filter result figure trIt is arranged to mean filter, obtains smooth transmissive rate figure mt:
H working group is arranged in (7a1), and each working group includes 256 threads, and each working group's alignment processing estimates transmission
The data line of rate figure t, wherein H is the height for estimating transmittance figure t；
(7a2) uses whole threads of i-th of working group, and by way of merging memory access, reading estimates transmittance figure t's
I-th row data, storage to local memory array g origin coordinates are l1In the contiguous memory space of=R+1, wherein R is filter
Radius；
The border extension pixel assignment of (7a3) to local memory array g:
First to the l of array g1A left margin extends pixel assignment, obtains pixel value of the array g at coordinate j:
G [j]=g [2l1- 1-j], wherein 0≤j < l1；
Again to the l of array g2A right margin extends pixel assignment, obtains array g in coordinate l1The pixel value at the place+W+k:
g[l1+ W+k]=g [l1+ W-1-k], in which: l2=(W+l1+R+M-1)/M×M-W-l1, W is to estimate transmittance figure t
Width, M be filter width M=2 × R+1,0≤k < l2；
(7a4) calculates block count: L=(W+R+l2)/M；
(7a5) with L thread preceding in i-th of working group, the prefix of the L piecemeal of parallel computation local memory array g and:
G [m]=g [m-1]+g [m], wherein g [m] is pixel value of the array g at coordinate m, to first of thread,
1+l × M≤m < 1+ (l+1) × M, l are thread number；
(7a6) calculates row mean filter result figure t with thread parallels all in i-th of working grouprIn the pixel of the i-th row
Value, then merge to form row mean filter result figure t with each row resultr, wherein trPixel value at coordinate position (i, j) are as follows:
tr(i, j)=(g [j+l1+R]+g[(j+M-1)/M×M]-g[j])/M；
(7a7) is according to row mean filter result figure tr, with ith thread, calculate the i-th filter result for arranging the 0th row:
Wherein, tr(j, i) is trIn the pixel value of coordinate position (j, i)；
(7a8) is according to row mean filter result figure tr, i-th arranges the 0th row filter result mt(0, i) is calculated with ith thread
The H-1 row filter result m of i-th columnt(j, i):
mt(j, i)=mt(j-1, i)+q (j+R, i)-q (j-R, i),
Wherein,
1≤j < H, mt(j-1, i) is the i-th column, -1 row filter result of jth；
The filter result m that (7a9) arranges the 0th row for i-thtThe H-1 row filter result m of (0, i) and the i-th columnt(j, i) carries out group
It closes, forms smooth transmissive rate figure mt。
The grayscale image I of (7b) to foggy imagegrayMean filter is carried out, smooth guide figure m is obtainedI；
(7c) is according to smooth transmissive rate figure mt, smooth guide figure mI, estimate the grayscale image of transmittance figure t and foggy image
Igray, obtain smooth scaled matrix ma, smoothed offset matrix mb:
(7c1) is according to the grayscale image I of foggy imagegrayWith smooth guide figure mI, calculate guidance figure variance matrix varI:
varI=fmean(Igray.*Igray,R)-mI.*mI,
Wherein fmean(Igray.*Igray, R) and it indicates to IgrayWith IgrayThe result of dot product carries out mean filter, and R is filter
Radius；
(7c2) is according to the grayscale image I for estimating transmittance figure, foggy imagegray, smooth guide figure mIWith smooth transmittance figure
mt, calculate guidance figure and estimate the covariance matrix cov of transmittance figureIt:
covIt=fmean(t.*Igray,R)-mt.*mI,
Wherein, fmean(t.*Igray, R) and it indicates to IgrayMean filter is carried out with the result of t dot product, R is filter half
Diameter .* are point multiplication operation symbol；
(7c3) is according to guidance figure variance matrix varI, guidance figure with estimate the covariance matrix cov of transmittance figureIt, it is smooth
Guidance figure mIWith smooth transmittance figure mt, calculate scaled matrix a, excursion matrix b:
B=mI-a.*mt
(7c4) carries out mean filter to scaled matrix a, excursion matrix b respectively, obtains smooth scaled matrix ma, it is smooth partially
Move matrix mb；
(7d) is according to smooth scaled matrix ma, smoothed offset matrix mbWith the grayscale image I of foggy imagegray, obtain fine saturating
Penetrate rate figure:
Step 8, defogging result figure is calculated.
(8a) calculates defogging result figure J pixel value J (z) at coordinate z:
Wherein,Respectively foggy image I, fine transmittance figurePixel value at coordinate z；
Pixel value J (z) at each coordinate is combined by (8b), forms defogging result figure J.
Step 9, it copies defogging result figure J to host memory, completes to the single frames defogging for having mist video.
The present invention realizes the video defogging of high speed by above step, can be subsequent video monitoring, image
The application such as identification provides clearly image source.
Effect of the invention can be further illustrated by following experiment:
Experiment 1, platform compatibility test:
The mist video resolution that has of this experiment test is 640 × 480, is calculated in different platform to before test video 500
Frame carries out defogging, exports average frame per second=frame number/processing time (second) of video, as shown in Table 1:
Table one, different platform video defogging are averaged frame per second table
As can be seen from Table I, the present invention is suitable for different platform, and the average frame per second of defogging output video is above now
There is solution.
Experiment 2, performance test
This experiment test equipment is AMD HD7750 video card, and choosing different resolution has a mist video, calculate to preceding 500 frame into
Row defogging exports the average frame per second of video, as shown in Table 2:
Table two, AMD HD7750 video card are averaged frame per second table to different resolution video defogging
As can be seen from Table II, the present invention has the processing speed per second more than 40 frames for high-definition picture, realizes
HD video is in real time handled.
Experiment 3, defog effect test
The mist video resolution that has of this experiment test is 640 × 480, and comparison has a frame image and the frame figure in mist video
As the clarity of image after defogging.
The 320th frame image of mist video has been taken, as shown in Fig. 2, defogging processing is carried out to the foggy image with the method for the present invention,
Image is as shown in Figure 3 after obtaining defogging.
2 Fig. 3 of comparison diagram shows that the present invention has good defog effect as it can be seen that image definition significantly improves after defogging.
To sum up, the present invention solves the defect of prior art, is suitable for a variety of parallel computing platforms, realizes high definition
Video is handled in real time, can be subsequent video monitoring, and the application such as image recognition provides clearly image source.
Claims (6)
1. the dark defogging parallel optimization method based on OpenCL, comprising:
(1) platform for supporting open computational language OpenCL is selected, which is initialized, and in therefrom choosing and capable of running
The equipment E of kernel function, the equipment include CPU, GPU, intel MIC accelerator card；
(2) there is a frame image I of mist video from the global memory that host memory copies equipment E to RGB color；
(3) gray value of each coordinate position of RGB color foggy image I is calculated by way of parallel processing, generates grayscale image
Igray, each coordinate position RGB triple channel minimum value is calculated by way of parallel processing, generates minimum value figure Imin；
(4) to minimum value figure IminMini-value filtering is carried out, dark channel diagram I is obtaineddark；
(5) according to dark channel diagram IdarkCalculate atmosphere light A:
(5a) calculates dark channel diagram IdarkHistogram h；
(5b) acquires the corresponding gray value lower bound l of maximum 0.1% pixel of gray value according to histogram h；
(5c) is threshold value to dark channel image I using ldarkIt carries out binaryzation and obtains bianry image I'；
The RGB triple channel figure I of (5d) to foggy image Ir、Ig、IbDot product bianry image I' respectively, obtains RGB triple channel grayscale image
As Ir'、Ig'、Ib'；
(5e) calculates separately RGB triple channel gray level image Ir'、Ig'、Ib' histogram hr、hg、hb；
(5f) passes through RGB three channel histogram hr、hg、hb, calculate atmosphere light A={ Ar,Ag,Ab, whereinI is the gray value in 0~255 variation, Ar、Ag、AbIt is respectively red, green, blue
The air light value of triple channel；
(6) according to the RGB triple channel figure I of atmosphere light A and foggy image Ir、Ig、Ib, calculate and estimate transmittance figure t；
(7) according to the grayscale image I of foggy imagegrayWith estimate transmittance figure t, obtain fine transmittance figure
(7a) is gone to transmittance figure t is estimated to mean filter, and row mean filter result figure t is obtainedr；Again to row mean filter
Result figure trIt is arranged to mean filter, obtains smooth transmissive rate figure mt:
H working group is arranged in (7a1), and each working group includes 256 threads, and each working group's alignment processing estimates transmittance figure
The data line of t, wherein H is the height for estimating transmittance figure t；
(7a2) uses whole threads of i-th of working group, and by way of merging memory access, reading estimates the i-th of transmittance figure t
Row data, storage to local memory array g origin coordinates are l1In the contiguous memory space of=R+1, wherein R is filter half
Diameter；
The border extension pixel assignment of (7a3) to local memory array g:
First to the l of array g1A left margin extends pixel assignment, obtains pixel value of the array g at coordinate j: g [j]=g [2l1-
1-j], wherein 0 <=j < l1；
Again to the l of array g2A right margin extends pixel assignment, obtains array g in coordinate l1The pixel value at the place+W+k: g [l1+W+
K]=g [l1+ W-1-k], wherein l2=(W+l1+R+M-1)/M×M-W-l1, W is the width for estimating transmittance figure t, and M is filtering
Device width M=2 × R+1,0 <=k < l2；
(7a4) calculates block count: L=(W+R+l2)/M；
(7a5) calculates the prefix and g [m] of the L piecemeal of local memory array g with L thread parallel preceding in i-th of working group
=g [m-1]+g [m], wherein g [m] is the prefix and array of one-row pixels, and each pair of point of the array final calculation result is answered
Be pixel value at m, to first of thread, 1+l × M <=m < 1+ (l+1) × M, l are thread number；
(7a6) calculates row mean filter result figure t with thread parallels all in i-th of working grouprIn the pixel value of the i-th row, then use
Each row result merges to form row mean filter result figure tr, wherein trPixel value at coordinate position (i, j) are as follows:
tr(i, j)=(g [j+l1+R]+g[(j+M-1)/M×M]-g[j])/M；
(7a7) is according to row mean filter result figure tr, with ith thread, calculate the i-th filter result m for arranging the 0th rowt(0, i):
Wherein, tr(j, i) is trIn the pixel value of coordinate position (j, i)；
(7a8) is according to row mean filter result figure tr, i-th arranges the 0th row filter result mt(0, i) calculates i-th with ith thread
The H-1 row filter result m of columnt(j, i):
mt(j, i)=mt(j-1,i)+q(j+R,i)-q(j-R,i)
Wherein,
1 <=j < H, mt(j-1, i) is the i-th column, -1 row filter result of jth；
The filter result m that (7a9) arranges the 0th row for i-tht(0, i) is placed on the H-1 row filter result m of the column of matrix i-thtBefore (j, i),
Form smooth transmissive rate figure mt；
The grayscale image I of (7b) to foggy imagegrayMean filter is carried out, smooth guide figure m is obtainedI；
(7c) is according to smooth transmissive rate figure mt, smooth guide figure mI, estimate the grayscale image I of transmittance figure t and foggy imagegray, obtain
To smooth scaled matrix ma, smoothed offset matrix mb；
(7d) is according to smooth scaled matrix ma, smoothed offset matrix mbWith the grayscale image I of foggy imagegray, obtain fine transmissivity
Figure
(8) according to RGB triple channel foggy image I, fine transmittance figureWith atmosphere light A, defogging result figure J is calculated；
(9) it copies defogging result figure J to host memory, completes to the single frames defogging for having mist video.
2. the dark defogging parallel optimization method according to claim 1 based on OpenCL, it is characterised in that: step
(3) parallel processing manner refers to single thread process single pixel point, different threads alignment processing image difference coordinate position in
Pixel.
3. the dark defogging parallel optimization method according to claim 1 based on OpenCL, it is characterised in that: step
(6) according to the RGB triple channel figure I of atmosphere light A and foggy image I inr、Ig、Ib, calculate and estimate transmittance figure t, as follows
It carries out:
Pixel value t (z) of the transmittance figure t at coordinate z is estimated in (6a) calculating:
Wherein Ir(y)、Ig(y)、IbIt (y) is respectively Ir、Ig、IbGray value at coordinate y, Ω (z) are centered on coordinate z
Square neighborhood；
Pixel value t (z) at each coordinate is combined by (6b), and transmittance figure t is estimated in formation.
4. the dark defogging parallel optimization method according to claim 1 based on OpenCL, it is characterised in that: step
Merging memory access in (7a2) refers to that, using the continuous thread of number in working group, the address accessed in global memory is continuous
Region of memory.
5. the dark defogging parallel optimization method according to claim 1 based on OpenCL, it is characterised in that: step
(7c) is carried out as follows:
(7c1) is according to the grayscale image I of foggy imagegray, smooth guide figure mI, calculate guidance figure variance matrix varI:
varI=fmean(Igray.*Igray,R)-mI.*mI,
Wherein fmean(Igray.*Igray, R) and it indicates to IgrayWith IgrayThe result of dot product carries out mean filter, and R is filter radius；
(7c2) basis estimates transmittance figure t, the grayscale image I of foggy imagegray, smooth guide figure mI, smooth transmissive rate figure mt, meter
It calculates guidance figure and estimates the covariance matrix cov of transmittance figureIt:
covIt=fmean(t.*Igray,R)-mt.*mI
Wherein, fmean(t.*Igray, R) and it indicates to IgrayMean filter is carried out with the result of t dot product, R is filter radius, and .* is
Point multiplication operation symbol；
(7c3) is according to guidance figure variance matrix varI, guidance figure and the covariance matrix cov for estimating transmittance figureIt, smooth guide
Scheme mI, smooth transmissive rate figure mt, calculate scaled matrix a, excursion matrix b:
(7c4) carries out mean filter to scaled matrix a, excursion matrix b respectively, obtains smooth scaled matrix ma, smoothed offset matrix
mb。
6. the dark defogging parallel optimization method according to claim 1 based on OpenCL, it is characterised in that: step
(8) according to RGB triple channel foggy image I, fine transmittance figure inWith atmosphere light A, defogging result figure J is calculated, as follows
It carries out:
(8a) calculates defogging result figure J pixel value J (z) at coordinate z:
Wherein,Respectively foggy image I, fine transmittance figurePixel value at coordinate z；
Pixel value J (z) at each coordinate is combined by (8b), forms defogging result figure J.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610808183.9A CN106469440B (en) | 2016-09-08 | 2016-09-08 | Dark defogging parallel optimization method based on OpenCL |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610808183.9A CN106469440B (en) | 2016-09-08 | 2016-09-08 | Dark defogging parallel optimization method based on OpenCL |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106469440A CN106469440A (en) | 2017-03-01 |
| CN106469440B true CN106469440B (en) | 2019-07-16 |
Family
ID=58230664
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610808183.9A Active CN106469440B (en) | 2016-09-08 | 2016-09-08 | Dark defogging parallel optimization method based on OpenCL |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106469440B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107424132B (en) * | 2017-07-25 | 2020-07-07 | 西安电子科技大学 | Optimization method for rapid image defogging |
| CN108093175B (en) * | 2017-12-25 | 2018-11-13 | 北京航空航天大学 | A kind of adaptive defogging method of real-time high-definition video and device |
| CN108765302B (en) * | 2018-03-29 | 2022-07-22 | 西安电子科技大学 | Image real-time defogging method based on GPU |
| CN109901840B (en) * | 2019-02-14 | 2020-10-27 | 中国科学院计算技术研究所 | Heterogeneous compilation optimization method for inter-thread redundancy deletion |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104751421A (en) * | 2015-03-10 | 2015-07-01 | 西安理工大学 | Method for achieving image defogging on FPGA |
| CN105488769A (en) * | 2015-12-08 | 2016-04-13 | 中国航空工业集团公司西安航空计算技术研究所 | Real time video defogging method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8340461B2 (en) * | 2010-02-01 | 2012-12-25 | Microsoft Corporation | Single image haze removal using dark channel priors |
-
2016
- 2016-09-08 CN CN201610808183.9A patent/CN106469440B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104751421A (en) * | 2015-03-10 | 2015-07-01 | 西安理工大学 | Method for achieving image defogging on FPGA |
| CN105488769A (en) * | 2015-12-08 | 2016-04-13 | 中国航空工业集团公司西安航空计算技术研究所 | Real time video defogging method |
Non-Patent Citations (4)
| Title |
|---|
| An improved sharpening algorithm for foggy picture based on dark-channel prior;Liu, H 等;《PROCEEDINGS OF THE 2015 INTERNATIONAL INDUSTRIAL INFORMATICS AND COMPUTER ENGINEERING CONFERENCE》;20150111;第2099-2104页 |
| 单幅雾天降质图像色彩复原算法研究;楚君;《中国优秀硕士学位论文全文数据库信息科技辑》;20151231;论文第3.3.3节，4.3节 |
| 基于 OpenCL 的图像去雾算法研究;刘晓光;《中国优秀硕士学位论文全文数据库信息科技辑》;20140329;论文第1.3，1.4，3.4，4.2，4.3，4.4节 |
| 高效均值滤波算法;王科俊等;《计算机应用研究》;20100228;论文引言 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106469440A (en) | 2017-03-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106469440B (en) | Dark defogging parallel optimization method based on OpenCL | |
| CN107301623B (en) | Traffic image defogging method and system based on dark channel and image segmentation | |
| CN102750674B (en) | Video image defogging method based on self-adapting allowance | |
| CN105354806B (en) | Rapid defogging method and system based on dark | |
| CN100539698C (en) | The video of interactive time-space unanimity is scratched drawing method in a kind of Digital Video Processing | |
| CN110634147B (en) | Image matting method based on bilateral guide up-sampling | |
| CN111914698B (en) | Human body segmentation method, segmentation system, electronic equipment and storage medium in image | |
| CN106023091B (en) | The real-time defogging method of image based on graphics processor | |
| CN108564549A (en) | A kind of image defogging method based on multiple dimensioned dense connection network | |
| CN103996170A (en) | Image edge saw-tooth eliminating method with super resolution | |
| CN105976337B (en) | A kind of image defogging method based on intermediate value guiding filtering | |
| CN105654440B (en) | Quick single image defogging algorithm based on regression model and system | |
| CN111899295B (en) | Monocular scene depth prediction method based on deep learning | |
| CN106933579A (en) | Image rapid defogging method based on CPU+FPGA | |
| Ling et al. | Single image dehazing using saturation line prior | |
| CN104021527B (en) | Rain and snow removal method in image | |
| CN110570381B (en) | Semi-decoupling image decomposition dark light image enhancement method based on Gaussian total variation | |
| CN106780380A (en) | A kind of image defogging method and system | |
| US20070200864A1 (en) | Method and system for gathering per-frame image statistics while preserving resolution and runtime performance in a real-time visual simulation | |
| Fang et al. | Single image dehazing and denoising with variational method | |
| CN114926407A (en) | Steel surface defect detection system based on deep learning | |
| CN110189390A (en) | A kind of monocular vision SLAM method and system | |
| CN111414938B (en) | Target detection method for bubbles in plate heat exchanger | |
| CN104822030B (en) | A kind of squaring antidote of irregular video based on anamorphose | |
| CN108629750A (en) | A kind of night defogging method, terminal device and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |