CN109615591A - A kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU - Google Patents
A kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU Download PDFInfo
- Publication number
- CN109615591A CN109615591A CN201811426126.XA CN201811426126A CN109615591A CN 109615591 A CN109615591 A CN 109615591A CN 201811426126 A CN201811426126 A CN 201811426126A CN 109615591 A CN109615591 A CN 109615591A
- Authority
- CN
- China
- Prior art keywords
- dimensional
- block
- thread
- image
- image block
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 45
- 230000009467 reduction Effects 0.000 claims abstract description 49
- 238000003860 storage Methods 0.000 claims abstract description 37
- 239000011159 matrix material Substances 0.000 claims abstract description 26
- 230000015654 memory Effects 0.000 claims abstract description 26
- 230000001133 acceleration Effects 0.000 claims abstract description 22
- 230000006870 function Effects 0.000 claims abstract description 21
- 230000009466 transformation Effects 0.000 claims description 40
- 238000001914 filtration Methods 0.000 claims description 28
- 230000008569 process Effects 0.000 claims description 20
- 238000005457 optimization Methods 0.000 claims description 12
- 238000001514 detection method Methods 0.000 claims description 7
- 238000009826 distribution Methods 0.000 claims description 7
- 230000010354 integration Effects 0.000 claims description 6
- 238000011524 similarity measure Methods 0.000 claims description 6
- 230000001174 ascending effect Effects 0.000 claims description 2
- 230000008859 change Effects 0.000 claims 1
- 238000010586 diagram Methods 0.000 description 10
- 210000003128 head Anatomy 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 241001269238 Data Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000011068 loading method Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 210000000214 mouth Anatomy 0.000 description 1
- 238000011946 reduction process Methods 0.000 description 1
- 230000000007 visual effect 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
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/10—Image enhancement or restoration using non-spatial domain filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20048—Transform domain processing
- G06T2207/20052—Discrete cosine transform [DCT]
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to technical field of image processing, specifically disclose a kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU, including image to be processed is carried out the pretreatment of boundary symmetric extension；Pretreated image data is sent in the global storage of GPU；Multithreaded network grid is created, the mode of access and the acceleration strategy of shared memory recycled for multiple times are merged using global storage, acceleration processing is carried out to the grouping of similar image Block- matching；The first step noise reduction estimated data of three-dimensional similar matrix is obtained using the parallel acceleration strategy of hard -threshold collaboration filter kernel function；It is reference with first step noise reduction estimated data, the parallel acceleration strategy of joint wiener collaboration filter kernel function obtains second step noise reduction estimated data；Second step noise reduction estimated data is rejected into extended boundary pixel.The present invention can improve data access speed and reduce repeated accesses delay, moreover it is possible to effectively remove the noise in image, be conducive to the processing of large-size images real-time noise-reducing.
Description
Technical field
The present invention relates to technical field of image processing, specifically disclose a kind of three-dimensional Block- matching accelerated parallel based on GPU
Noise-reduction method.
Background technique
Digital picture is big due to being influenced usually to contain by imaging device and external environment in acquisition and transmission process
Noise is measured, picture quality is influenced.It is past in CT image due to being influenced by low dosage voltage and current especially in medical application
Toward a large amount of noise can be introduced, cause deteriroation of image quality, influences the clinical disease diagnosis of doctor.
Common three-dimensional Block- matching (BM3D) algorithm combines local, non local, multiple dimensioned sparse, adaptive by its own
Filtering noise reduction feature, it is considered to be best Image denoising algorithm at present, but the algorithm is based on the collaboration filter of similar image block
Wave, algorithm complexity is high, computationally intensive, when handling large-sized CT image data, takes a long time, treatment effeciency is lower, can not
Meet real work demand.
Therefore, it is necessary to a kind of methods that can solve the above problem.
Summary of the invention
In order to overcome shortcoming and defect existing in the prior art, the purpose of the present invention is to provide one kind to be based on GPU simultaneously
The three-dimensional Block- matching noise-reduction method that row accelerates.
To achieve the above object, the present invention uses following scheme.
A kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU, comprising:
Image to be processed is subjected to the pretreatment of boundary symmetric extension at the end CPU；
By pretreated image data from the global storage that CPU host side is sent to GPU；
Multithreaded network grid is created, the mode and shared memory recycled for multiple times of access are merged using global storage
Acceleration strategy, parallel acceleration processing is carried out to the grouping of similar image Block- matching；
Estimated using the first step noise reduction that the parallel acceleration strategy of hard -threshold collaboration filter kernel function obtains three-dimensional similar matrix
It counts；
It is reference with first step noise reduction estimated data, the parallel acceleration strategy of joint wiener collaboration filter kernel function obtains the
Two step noise reduction estimated datas；
Second step noise reduction estimated data is sent to CPU host side from GPU, and rejects extended boundary pixel to obtain noise reduction
Image afterwards.
Further, the creation multithreaded network grid, comprising:
With the image block matching process of each reference block of image for a thread block block, each detection figure in search window
As the similitude matching process of block and reference block is that per thread thread carries out thread distribution；
With the step-length of certain pixel respectively from ranks direction selection reference image block incremented by successively, and referred to according in image
Image block number determines the size of thread net grid, determines thread block according to the image block number of reference image block search window
Size.
Further, the mode and shared memory recycled for multiple times that access is merged using global storage
Acceleration strategy carries out parallel acceleration processing to the grouping of similar image Block- matching, comprising:
All thread thread in the same warp are executed into sequential cells in same instruction access global storage,
Merge access module to obtain；
By search window be divided into size be 32*32 4 piecemeals, and in each piecemeal with the thread of block (16,16) into
Row Similarity measures obtain similar block,Wherein distance of the d between image block is defined as two figures
As size of the mould divided by image block of block corresponding element numerical difference, τthreodFor the suitable distance threshold of selection；
By in search window in pel data Circulant Block shared memory, and be arranged threadIdx.x < 16 and
threaIdx.y<16；
The most like image block of defined amount is found using the parallel strategy of minimum value reduction.
Further, the strategy parallel using minimum value reduction finds the most like image block of defined amount, packet
It includes:
The similar image block of reference image block is gathered into three-dimensional matrice according to the ascending sequence of similarity distance；
It is corresponding to enable n distance value D [n] that n thread is obtained with Similarity measures respectively；
The value of ith thread is compared with the distance value of (i+n/2) a thread and smaller value is placed on left half
In, the larger value is placed on right half, then left part by stages is D [0] to D [n/2], and right part by stages is D [n/2] to D [n]；
After completing thread parallel relatively, Thread Count will be compared and halved, above-mentioned comparison is carried out to left half zone distance value, directly
Halve to multiple alternative line number of passes and compares rear left partial section value and be reduced to D [0]；
Using D [0] as the minimum value in distance value, D [n] starting access position is moved back one and repeated the above steps and is sought
Minimum value is looked for, until finding the similar image block apart from the smallest defined amount.
Further, described that three-dimensional similar matrix is obtained using the parallel acceleration strategy of hard -threshold collaboration filter kernel function
First step noise reduction estimated data, comprising:
Instruct hybrid optimization to accelerate, by the three-dimensional direct transform of three-dimensional similar matrix, hard -threshold filtering, three-dimensional inverse transformation and
The process integration of Weighted estimation is in hard -threshold collaboration filter kernel function；Wherein the three-dimensional direct transform includes successively carrying out two
Tie up Bi-orthogonal Spline Wavelet Transformation direct transform and one-dimensional Walsh-Hadanjard Transform；The three-dimensional inverse transformation is one-dimensional including successively carrying out
Walsh-Hadanjard Transform and two-dimentional Bi-orthogonal Spline Wavelet Transformation inverse transformation；
According to the size of reference image block, a certain number of similar image blocks are chosen, and keep thread grid grid constant
With the size of setting thread block block；
The similar image block number of selection is sent according to from global storage to altogether using the mode that global storage merges access
It enjoys in memory to constitute three-dimensional similar matrix；
Two-dimentional Bi-orthogonal Spline Wavelet Transformation direct transform and one-dimensional Walsh-Hadanjard Transform are carried out to three-dimensional matrice, and become
Progress hard -threshold filtering in domain is changed, passes through one-dimensional Walsh-Hadanjard Transform and two-dimentional bi-orthogonal spline after hard -threshold filtering again
Wavelet inverse transformation obtains the first step noise reduction estimated data of image block,
Gray value in hard -threshold filtered image block is weighted and averaged, weighted average is assigned to the single of image block
Pixel, and introduce triumphant plucked instrument window coefficient in weighted average and be weighted optimization to obtain first step image noise reduction value；
Meanwhile three-dimensional similar matrix being carried out to four filter coefficients lpd, hpd of two-dimentional Bi-orthogonal Spline Wavelet Transformation transformation,
Lpr, hpr are stored in constant storage, and define private variable storage intermediate result in each thread using register.
Further, it is described with first step noise reduction estimated data be reference, joint wiener collaboration filter kernel function it is parallel
Acceleration strategy obtains second step noise reduction estimated data, comprising:
It instructs hybrid optimization to accelerate, by the three-dimensional direct transform of three-dimensional similar matrix, Wiener filtering, three-dimensional inverse transformation and adds
The process integration of kernel estimators is in wiener collaboration filter kernel function；Wherein the three-dimensional direct transform include successively carry out two dimension from
Dissipate cosine direct transform and one-dimensional Walsh-Hadanjard Transform；The three-dimensional inverse transformation includes successively carrying out one-dimensional Walsh-hada
Hadamard transform and 2-D discrete cosine inverse transformation；
According to the size of reference image block, a certain number of similar image blocks are chosen, and keep thread grid grid constant
With the size of setting thread block block；
The similar image block number of selection is sent according to from global storage to altogether using the mode that global storage merges access
It enjoys in memory to constitute three-dimensional similar matrix；
With first step noise reduction estimated data be reference, by initial three-dimensional similar matrix carry out 2-D discrete cosine direct transform and
One-dimensional Walsh-Hadanjard Transform；Wiener filtering is carried out, and one-dimensional Wall is carried out to the three-dimensional similar matrix after Wiener filtering
Assorted-Hadamard transform and 2-D discrete cosine inverse transformation obtain second step noise reduction estimated data；
Gray value in image block after Wiener filtering is weighted and averaged, weighted average is assigned to the single picture of image block
Element, and introduce triumphant plucked instrument window coefficient in weighted average and be weighted optimization to obtain second step image noise reduction value；
Meanwhile Two Dimension Discrete Cosine is stored in private variable defined in per thread using register
In.
Further, when boundary symmetric extension pre-processes, the border column pixel of the left and right sides is carried out respectively first symmetrical
Then extension carries out symmetric extension to the border row pixel of upper and lower two sides respectively again, and the pixel width of border extension is by searching
The radius size of rope window determines.
Beneficial effects of the present invention: a kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU is provided, merging is passed through
The mode of access once reads data required for each GPU thread block in shared memory from global storage, and adopts
With the strategy of shared memory is recycled, to improve data access speed and reduce repeated accesses delay, significant increase
Algorithm overall performance, improves computational efficiency；Cooperate hard -threshold collaboration filter kernel function and wiener collaboration filter kernel simultaneously
Function can effectively remove the noise in image, be conducive to the processing of large-size images real-time noise-reducing.
Detailed description of the invention
Fig. 1 is the flow diagram of the embodiment of the present invention.
Fig. 2 is that image of the embodiment of the present invention carries out the pretreated schematic diagram of boundary symmetric extension.
Fig. 3 is the schematic diagram of thread of embodiment of the present invention grid distribution.
Fig. 4 is the schematic diagram of shared memory of embodiment of the present invention employment mechanism.
Fig. 5 is the schematic diagram of minimum value of embodiment of the present invention reduction sorting in parallel.
Fig. 6 is schematic diagram of the present invention using the original CT image of head mould.
Fig. 7 is the schematic diagram of the CT image of present invention denoising back mould.
Fig. 8 is schematic diagram of the present invention using the original CT image of body mould.
Fig. 9 is the schematic diagram of the CT image of body mould after the present invention denoises.
Specific embodiment
For the ease of the understanding of those skilled in the art, the present invention is made further below with reference to examples and drawings
Bright, the content that embodiment refers to not is limitation of the invention.
A kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU, as shown in Figure 1, comprising:
CT image boundarg pixel, which crosses the border, in order to prevent not can be carried out noise reduction process, and it is symmetrical that image to be processed is carried out boundary
Extension pretreatment.It needs when as shown in Fig. 2, carrying out symmetric extension pretreatment first to noise variance, similar block number, search window radius
Etc. parameters be configured, and the pixel wide of border extension is determined by search window radius size, and general value is 16 pixels.
After completing the pretreatment of image boundary symmetric extension, pretreated image data will be extended and be sent to from host side
In the global storage of GPU；
Thread distribution need to be carried out when creating multithreaded network grid, as shown in figure 3, considering the similar Block- matching of different reference blocks
The irrelevant property of process, it is every in search window with the image block matching process of each reference block of image for a thread block block
The similitude matching process of a detection image block and reference block is per thread thread.Then in the picture along the line of the column direction with
The step-length of every 3 pixels selection reference image block incremented by successively, when reference image block number is (M*N) in image, then thread
The size of grid grid is (M, N), and the 32*32 image block detected in reference block search window carries out similitude matching, thread block
The constant magnitude of block is set as (32,32).
Then, the mode of access and the acceleration strategy of shared memory recycled for multiple times are merged using global storage,
Parallel acceleration processing is carried out to the grouping of similar image Block- matching.As shown in figure 4, making all thread thread in the same warp
When being carried out continuous unit in same instruction access global storage, best access module is obtained.One of them
Warp beam has 32 threads, therefore search window is divided into 4 piecemeals that size is 32*32, it is made to meet global storage merging
The requirement of Access Optimization reaches global storage bandwidth communication peak value as far as possible.And in each piecemeal with block (16,
16) the thread method of salary distribution carries out parallel Similarity measures and obtains similar block.
When carrying out Similarity measures, with reference image block PkCentered on, radius is sliding detection pixel-by-pixel in the search window of R
Window (size N*N) calculates the image block that judgement detection window is chosenWith reference image block PkSimilitude, and by similar figure
As block is assembled to obtain reference image block PkThree-dimensional similar image block group.Similitude judgment formula are as follows:
In formula, distance of the d between image block is defined as the mould of two image block corresponding element numerical differences divided by image block
Size, τthreodFor the suitable distance threshold of selection, meet d < τthreodWhen, that is, think that two image blocks are similar, conversely, then
Not so.
After completing the operation of similitude PARALLEL MATCHING, block data is copied to shared memory.Since global storage is every
It is secondary to access the delay for having up to 400~600 clock cycle, and there is a large amount of again for pel data in image block matching process
Repeated accesses, therefore reduced as far as possible during copying the pel data in search window the shared memory of cache to
To the access times of global storage.And limited by GPU hardware condition, the maximum thread of thread block block is 1024, will
When the pel data of 32*32 detection block disposably copies shared memory to from global storage in search window, pixel is total
Number has exceeded thread maximum number, in order to effectively solve this problem, copies using by pel data Circulant Block in search window
Strategy into shared memory is accelerated parallel.Simultaneously in order to avoid being overlapped as piecemeal caused by image block repeated matching
ThreadIdx.x < 16 and threaIdx.y < 16 are arranged after block data is copied to shared memory in situation.
The most like image block that defined amount is found using the parallel strategy of minimum value reduction, as shown in figure 5, according to reality
The most like image block that situation chooses certain amount participates in subsequent processing, i.e., distance need to be selected in similar image block most
Small similar image block.It sorts from small to large according to similarity distance referring initially to the similar image block of image block and is gathered into three-dimensional square
It is corresponding to enable n distance value D [n] that n thread is obtained with Similarity measures respectively for battle array；By the value of ith thread and (i+
N/2) distance value of a thread is compared and smaller value is placed in left half, and the larger value is placed on right half, then left part subregion
Between be D [0] to D [n/2], right part by stages be D [n/2] to D [n]；At this time n/2 thread parallel relatively after, minimum value is certain
It is D [0] to D [n/2] in left part by stages.After completing n/2 thread parallel relatively, compares Thread Count and halve into n/4, to left half
Section equally carries out above-mentioned comparison, then the left part by stages where minimum value is also reduced into D [0] to D [n/4]；Repeat above-mentioned step
Suddenly, until left part by stages is reduced into D [0], then D [0] is the minimum value in all distance values at this time.Each reduction sorting in parallel
After finding minimum value, one will be moved back apart from the starting of array D access position, array length becomes n-1, then in circulating repetition
The most like image block of defined amount can be found by stating reduction operations process finally, until finding apart from the smallest defined amount
Similar image block.Compare Thread Count with this each reduction and be all reduced into original half, greatly reduces the time of sequence.
Either Floating-point Computation instruction, load instruction or branch instruction occupy instruction processing bandwidth, the finger of each SM
Processing bandwidth is enabled all to be limited.Therefore accelerated using instruction hybrid optimization, by the three-dimensional direct transform (two dimension of three-dimensional similar matrix
Bi-orthogonal Spline Wavelet Transformation direct transform+one-dimensional Walsh-Hadanjard Transform), hard -threshold filtering, three-dimensional inverse transformation (one-dimensional Walsh-
Hadamard transform+two dimension Bi-orthogonal Spline Wavelet Transformation inverse transformation) and the process integration of Weighted estimation unified cooperateed with to a hard -threshold
In filter kernel function, to reduce reverse cyclic loadings, the copy instruction of unnecessary intermediate variable, unnecessary time-consuming is saved.
Thread distribution is carried out, the size of reference image block is 8*8, and hard -threshold collaboration filtering selects 16 similar image block ginsengs
With processing, therefore the size of thread block block is specifically configured to (64,16), thread grid grid is remained unchanged, and is equally used
The mode for merging access copies the most like image block data of selection in shared memory to constitute three from global storage to
Tie up similar matrix
Three-dimensional matriceProgress two-dimentional Bi-orthogonal Spline Wavelet Transformation (BIOR) one-dimensional Walsh-Hadanjard Transform between direct transform and block,
And hard -threshold filtering is carried out in the transform domain as illustrated, pass through one-dimensional Walsh-Hadanjard Transform and two-dimentional Bi-orthogonal Spline Wavelet Transformation after filtering again
(BIOR) inverse transformation obtains the first step noise reduction estimated value of all image blocks in groupProcess is
WhereinFor the one-dimensional Walsh-Hadanjard Transform of radial direction between the two-dimentional BIOR transformation of image block each in group and block；
The two-dimentional BIOR of radial one-dimensional Walsh-Hadanjard Transform and each image block is inverse between the block of three-dimensional matrice after hard -threshold filtering
Transformation；γ is hard -threshold filtering processing；
After three-dimensional direct transform, noise is often focused at coefficient in transform domain smaller value, and true detail information concentrates on becoming
It changes at domain coefficient the larger value, therefore the coefficient in transform domain for being less than threshold parameter is set 0 by hard -threshold filtering, other coefficients retain not
Become, process is shown below:
In formula, x is three-dimensional matriceTransformation coefficient after three-dimensional direct transform,For the shrinkage parameters of hard -threshold filtering, σ
For the noise bias of estimation.
After hard -threshold collaboration filtering, each pixel of each image block obtains an estimated value, but for a certain
A pixel i, is likely to appear in multiple images block, thus possess multiple estimated values, the picture for needing to have these block of overlapping
First estimated value is weighted and averaged to obtain the basic noise reduction estimated value of pixel i, basic weightIt is filtered by hard -threshold
Three-dimensional matrice afterwardsIn non-zero transform domain coefficients numberIt determines,
In weighted mean procedure, in order to be further reduced edge effect, triumphant plucked instrument window coefficient W is addedkaiserIt is weighted poly-
Collection, then basic noise reduction estimates the value of any pixel i in image are as follows:
In formula,
Simultaneously as small echo becomes when the two-dimentional Bi-orthogonal Spline Wavelet Transformation for carrying out image block to three-dimensional similar matrix converts
Four the filter coefficients lpd, hpd, lpr, hpr changed are constants, and need frequently accessed repeatedly by per thread, therefore by its
It is stored in and possesses in the constant storage that caching accelerates, register is made full use of to accelerate access speed, save operation time.Deposit
Device and shared memory are located at GPU chip interior, are two most fast memories of access speed respectively.Two-dimentional bi-orthogonal spline is small
Wave conversion process is related to interative computation, needs defined variable storage intermediate conversion as a result, register is made full use of simultaneously, each
Private variable storage intermediate result is defined in thread, to improve interative computation efficiency.
It is reference with first step noise reduction estimation image, the estimation of Wiener filtering noise reduction is carried out to original noisy image.It is same first
Sample is accelerated using instruction hybrid optimization, by three-dimensional direct transform (the 2-D discrete cosine direct transform+one-dimensional Wall of three-dimensional similar matrix
Assorted-Hadamard transform), Wiener filtering, three-dimensional inverse transformation (one-dimensional Walsh-Hadanjard Transform+2-D discrete cosine inverse transformation)
And the process integration of Weighted estimation is unified into wiener collaboration filter kernel function.
When carrying out thread distribution optimization, Wiener filtering selects 32 similar image block participations processing, therefore by thread block block
Size be specifically configured to (64,32), thread grid grid is remained unchanged, and equally merge access copy shared memory to,
Constitute new three-dimensional similar matrixThe similar block matrix of three-dimensional of original image is obtained from original noisy image simultaneouslyIt is right
Two three-dimensional matricesWith2-D discrete cosine direct transform (DCT) and one-dimensional Walsh-Hadanjard Transform are carried out respectively,
Then experience wiener coefficient is calculated with the three-dimensional direct transform matrix of basis estimation image
Followed by experience wiener coefficient to the three-dimensional matrice of original noisy imageWiener filtering processing is carried out, has been handled
Cheng Houzai is estimated by the noise reduction that all image blocks in group can be obtained in inverse transformation
Principle using the similar first step is estimated to be weighted and averaged to superposition image member, the difference is that wherein basic weightBy
The decision of experience wiener coefficient,Then second step noise reduction estimates the value of any pixel i in image are as follows:
Simultaneously as transformation coefficient is related to the great trigonometric function of expense in the two-dimension discrete cosine transform of image block
It calculates, therefore is precalculated out, and register is made full use of to store it in the private variable that per thread defines
In.Second step noise reduction estimated data is then sent to CPU host side from GPU, and rejects extended boundary pixel to obtain noise reduction
Image afterwards.
A kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU provided by the invention, it is contemplated that the figure of cross-thread
As there is repetitions of a large amount of data access, and each of global storage is accessed when having up to 400~600 in block detection
The delay in clock period once reads shared memory from global storage using by data required for each GPU thread block,
And the strategy of shared memory is reused, to save repeated accesses delay, promote computational efficiency.And data are stored from the overall situation
Device is all made of the mode for merging access when reading shared memory, can improve the access speed of data, reach complete as far as possible
The peak value of office's bandwidth of memory, can will carry out GPU acceleration based on three-dimensional block matching algorithm, significant increase algorithm overall performance,
The processing time is saved, calculating speed improves nearly 80 times than conventional serial algorithm；Hard -threshold is cooperated to cooperate with filter kernel simultaneously
Function and wiener cooperate with filter kernel function, have not only effectively removed the noise in image, also help and large scale CT is schemed
The real-time noise-reducing of picture is handled.
More specifically, it includes noise original CT image that Fig. 6, which is the head mould that the present invention uses, scanned position is the oral cavity of human body
Tooth, Fig. 7 are using the head mould CT image after the method for the present invention denoising；Fig. 8 is that the present invention is to include much noise using body mould
Original CT image, Fig. 9 be using the method for the present invention denoising after body mould CT image.Pass through 6,7 and Fig. 8 of comparison diagram, 9 experiment
As a result, from can be seen that the method for the present invention has effectively removed the noise in original image, image entirety clarity on improvement of visual effect
Preferably, edge detail information is also kept good.From processing speed, original serial algorithm process breadth is 2048*2048 size
CT image need 1.5 hours, and the method for the present invention processing identical image time-consuming then only need 69 seconds, speed improve by
Nearly 80 times.
The above is only a preferred embodiment of the present invention, for those of ordinary skill in the art, according to the present invention
Thought, there will be changes in the specific implementation manner and application range, and the content of the present specification should not be construed as to the present invention
Limitation.
Claims (7)
1. a kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU characterized by comprising
Image to be processed is subjected to the pretreatment of boundary symmetric extension at the end CPU；
By pretreated image data from the global storage that CPU host side is sent to GPU；
Multithreaded network grid is created, the mode of access and adding for shared memory recycled for multiple times are merged using global storage
Fast strategy carries out parallel acceleration processing to the grouping of similar image Block- matching；
The first step noise reduction estimative figure of three-dimensional similar matrix is obtained using the parallel acceleration strategy of hard -threshold collaboration filter kernel function
According to；
It is reference with first step noise reduction estimated data, the parallel acceleration strategy of joint wiener collaboration filter kernel function obtains second step
Noise reduction estimated data；
Second step noise reduction estimated data is sent to CPU host side from GPU, and after rejecting extended boundary pixel to obtain noise reduction
Image.
2. a kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU according to claim 1, which is characterized in that
The creation multithreaded network grid, comprising:
With the image block matching process of each reference block of image for a thread block block, each detection image block in search window
Similitude matching process with reference block is that per thread thread carries out thread distribution；
With the step-length of certain pixel respectively from ranks direction selection reference image block incremented by successively, and according to reference picture in image
Block number determines the size of thread net grid, determines that thread block's is big according to the image block number of reference image block search window
It is small.
3. according to a kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU described in claim 1, which is characterized in that institute
The acceleration strategy for stating the mode and shared memory recycled for multiple times that merge access using global storage, to similar image block
Matched packet carries out parallel acceleration processing, comprising:
All thread thread in the same warp are executed into sequential cells in same instruction access global storage, to obtain
Access module must be merged；
Search window is divided into 4 piecemeals that size is 32*32, and phase is carried out with the thread of block (16,16) in each piecemeal
It is calculated like property and obtains similar block,Wherein distance of the d between image block is defined as two image blocks
Size of the mould of corresponding element numerical difference divided by image block, τthreodFor the suitable distance threshold of selection；
By in search window in pel data Circulant Block shared memory, and be arranged threadIdx.x < 16 and threaIdx.y <
16；
The most like image block of defined amount is found using the parallel strategy of minimum value reduction.
4. according to a kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU as claimed in claim 3, which is characterized in that institute
State the most like image block that defined amount is found using the parallel strategy of minimum value reduction, comprising:
The similar image block of reference image block is gathered into three-dimensional matrice according to the ascending sequence of similarity distance；
It is corresponding to enable n distance value D [n] that n thread is obtained with Similarity measures respectively；
The value of ith thread is compared with the distance value of (i+n/2) a thread and smaller value is placed in left half, compared with
Big value is placed on right half, then left part by stages is D [0] to D [n/2], and right part by stages is D [n/2] to D [n]；
After completing thread parallel relatively, Thread Count will be compared and halved, above-mentioned comparison is carried out to left half zone distance value, until more
Secondary relatively Thread Count, which halves, to be compared rear left partial section value and is reduced to D [0]；
Using D [0] as the minimum value in distance value, D [n] starting access position is moved back one and repeats the above steps searching most
Small value, until finding the similar image block apart from the smallest defined amount.
5. a kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU according to claim 1, which is characterized in that
The first step noise reduction estimative figure that three-dimensional similar matrix is obtained using the parallel acceleration strategy of hard -threshold collaboration filter kernel function
According to, comprising:
Hybrid optimization is instructed to accelerate, by the three-dimensional direct transform of three-dimensional similar matrix, hard -threshold filtering, three-dimensional inverse transformation and weighting
The process integration of estimation is in hard -threshold collaboration filter kernel function；Wherein the three-dimensional direct transform includes successively carrying out two dimension pair
Orthogonal Spline Wavelet direct transform and one-dimensional Walsh-Hadanjard Transform；The three-dimensional inverse transformation includes successively carrying out one-dimensional Wall
Assorted-Hadamard transform and two-dimentional Bi-orthogonal Spline Wavelet Transformation inverse transformation；
According to the size of reference image block, a certain number of similar image blocks are chosen, and keeps thread grid grid constant and sets
Set the size of thread block block；
The similar image block number evidence of selection is sent from global storage to shared using the mode that global storage merges access and is deposited
To constitute three-dimensional similar matrix in reservoir；
Two-dimentional Bi-orthogonal Spline Wavelet Transformation direct transform and one-dimensional Walsh-Hadanjard Transform are carried out to three-dimensional matrice, and in transform domain
Middle progress hard -threshold filtering passes through one-dimensional Walsh-Hadanjard Transform and two-dimentional Bi-orthogonal Spline Wavelet Transformation again after hard -threshold filtering
Inverse transformation obtains the first step noise reduction estimated data of image block,
Gray value in hard -threshold filtered image block is weighted and averaged, weighted average is assigned to the single picture of image block
Element, and introduce triumphant plucked instrument window coefficient in weighted average and be weighted optimization to obtain first step image noise reduction value；
Meanwhile three-dimensional similar matrix being carried out to four filter coefficients lpd, hpd, lpr of two-dimentional Bi-orthogonal Spline Wavelet Transformation transformation,
Hpr is stored in constant storage, and defines private variable storage intermediate result in each thread using register.
6. a kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU according to claim 1, which is characterized in that
Described with first step noise reduction estimated data is reference, and the parallel acceleration strategy of joint wiener collaboration filter kernel function obtains second step
Noise reduction estimated data, comprising:
It instructs hybrid optimization to accelerate, the three-dimensional direct transform of three-dimensional similar matrix, Wiener filtering, three-dimensional inverse transformation and weighting is estimated
The process integration of meter is in wiener collaboration filter kernel function；Wherein the three-dimensional direct transform includes successively carrying out more than two-dimensional discrete
String direct transform and one-dimensional Walsh-Hadanjard Transform；The three-dimensional inverse transformation includes successively carrying out one-dimensional Walsh-Hadamard change
It changes and 2-D discrete cosine inverse transformation；
According to the size of reference image block, a certain number of similar image blocks are chosen, and keeps thread grid grid constant and sets
Set the size of thread block block；
The similar image block number evidence of selection is sent from global storage to shared using the mode that global storage merges access and is deposited
To constitute three-dimensional similar matrix in reservoir；
It is reference with first step noise reduction estimated data, initial three-dimensional similar matrix is subjected to 2-D discrete cosine direct transform and one-dimensional
Walsh-Hadanjard Transform；Wiener filtering is carried out, and one-dimensional Walsh-Kazakhstan is carried out to the three-dimensional similar matrix after Wiener filtering
Second step noise reduction estimated data is obtained up to Hadamard transform and 2-D discrete cosine inverse transformation；
Gray value in image block after Wiener filtering is weighted and averaged, weighted average is assigned to the single pixel of image block,
And triumphant plucked instrument window coefficient is introduced in weighted average and is weighted optimization to obtain second step image noise reduction value；
Meanwhile Two Dimension Discrete Cosine being stored in private variable defined in per thread using register.
7. a kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU according to claim 1, which is characterized in that
When boundary symmetric extension pre-processes, symmetric extension is carried out to the border column pixel of the left and right sides respectively first, it is then right respectively again
The border row pixel of upper and lower two sides carries out symmetric extension, and the pixel width of border extension is true by the radius size of search window
It is fixed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811426126.XA CN109615591A (en) | 2018-11-27 | 2018-11-27 | A kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811426126.XA CN109615591A (en) | 2018-11-27 | 2018-11-27 | A kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109615591A true CN109615591A (en) | 2019-04-12 |
Family
ID=66005289
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811426126.XA Pending CN109615591A (en) | 2018-11-27 | 2018-11-27 | A kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109615591A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111179201A (en) * | 2019-12-31 | 2020-05-19 | 广州市百果园信息技术有限公司 | Video denoising method and electronic equipment |
| CN117074443A (en) * | 2023-10-17 | 2023-11-17 | 广东天信电力工程检测有限公司 | X-ray nondestructive testing robot for power transmission line |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110115806A1 (en) * | 2009-11-19 | 2011-05-19 | Rogers Douglas H | High-compression texture mapping |
| CN102547289A (en) * | 2012-01-17 | 2012-07-04 | 西安电子科技大学 | Fast motion estimation method realized based on GPU (Graphics Processing Unit) parallel |
| CN107369169A (en) * | 2017-06-08 | 2017-11-21 | 温州大学 | The approximate most like image block matching method transmitted based on direction alignment with matching that a kind of GPU accelerates |
-
2018
- 2018-11-27 CN CN201811426126.XA patent/CN109615591A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110115806A1 (en) * | 2009-11-19 | 2011-05-19 | Rogers Douglas H | High-compression texture mapping |
| CN102547289A (en) * | 2012-01-17 | 2012-07-04 | 西安电子科技大学 | Fast motion estimation method realized based on GPU (Graphics Processing Unit) parallel |
| CN107369169A (en) * | 2017-06-08 | 2017-11-21 | 温州大学 | The approximate most like image block matching method transmitted based on direction alignment with matching that a kind of GPU accelerates |
Non-Patent Citations (1)
| Title |
|---|
| 袁龙杰: "基于GPU的三维块匹配去噪并行算法研究", 《NSTL国家科技图书文献中心》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111179201A (en) * | 2019-12-31 | 2020-05-19 | 广州市百果园信息技术有限公司 | Video denoising method and electronic equipment |
| CN111179201B (en) * | 2019-12-31 | 2023-04-11 | 广州市百果园信息技术有限公司 | Video denoising method and electronic equipment |
| CN117074443A (en) * | 2023-10-17 | 2023-11-17 | 广东天信电力工程检测有限公司 | X-ray nondestructive testing robot for power transmission line |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Gómez-Ríos et al. | Towards highly accurate coral texture images classification using deep convolutional neural networks and data augmentation | |
| Vijay et al. | Automated brain tumor segmentation and detection in MRI using enhanced Darwinian particle swarm optimization (EDPSO) | |
| CN110782477A (en) | Moving target rapid detection method based on sequence image and computer vision system | |
| CN109615591A (en) | A kind of three-dimensional Block- matching noise-reduction method accelerated parallel based on GPU | |
| CN110517195A (en) | Unsupervised SAR image denoising method | |
| CN112102217B (en) | Method and system for quickly fusing visible light image and infrared image | |
| CN111105452A (en) | High-low resolution fusion stereo matching method based on binocular vision | |
| CN110866870A (en) | Super-resolution processing method for amplifying medical image by any multiple | |
| CN107085826B (en) | Image super-resolution reconstruction method based on weighted overlap non-local regression prior | |
| Wang et al. | Multi-wavelet residual dense convolutional neural network for image denoising | |
| CN111931927B (en) | Method and device for reducing occupation of computing resources in NPU | |
| Hsiao et al. | A parameterizable digital-approximated 2D Gaussian smoothing filter for edge detection in noisy image | |
| CN116503270A (en) | Electronic tomographic image denoising method based on block matching three-dimensional collaborative filtering | |
| CN113936117B (en) | High-frequency region enhanced luminosity three-dimensional reconstruction method based on deep learning | |
| CN115631223A (en) | Multi-view stereo reconstruction method based on self-adaptive learning and aggregation | |
| CN109840892A (en) | A kind of infrared video Enhancement Method based on High Level Synthesis | |
| Peng et al. | Low-light image enhancement based on FPGA | |
| CN113793280A (en) | Real image noise reduction method combining local noise variance estimation and BM3D block matching | |
| Morris et al. | Breaking the ‘Ton’: achieving 1% depth accuracy from stereo in real time | |
| CN112465719A (en) | Transform domain image denoising method and system | |
| Kryjak et al. | Pipeline implementation of peer group filtering in FPGA | |
| Zhang et al. | Non-blind image deconvolution using deep dual-pathway rectifier neural network | |
| Lew et al. | Adaptive Gaussian Wiener Filter for CT-Scan Images with Gaussian Noise Variance | |
| WO2019178717A1 (en) | Binocular matching method, visual imaging device and device with storage function | |
| Huang et al. | Fluorescence image denoising based on self-supervised deep learning |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190412 |
|
| RJ01 | Rejection of invention patent application after publication |