CN107909558A - A kind of non-local mean image de-noising method based on unsupervised learning - Google Patents
A kind of non-local mean image de-noising method based on unsupervised learning Download PDFInfo
- Publication number
- CN107909558A CN107909558A CN201711312994.0A CN201711312994A CN107909558A CN 107909558 A CN107909558 A CN 107909558A CN 201711312994 A CN201711312994 A CN 201711312994A CN 107909558 A CN107909558 A CN 107909558A
- Authority
- CN
- China
- Prior art keywords
- image
- mrow
- msub
- window
- pixel
- 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 77
- 230000000694 effects Effects 0.000 claims abstract description 26
- 238000000205 computational method Methods 0.000 claims description 6
- 238000004422 calculation algorithm Methods 0.000 abstract description 12
- 230000008901 benefit Effects 0.000 abstract description 5
- 244000144992 flock Species 0.000 abstract description 2
- 230000009467 reduction Effects 0.000 abstract description 2
- 238000001914 filtration Methods 0.000 description 28
- 238000010586 diagram Methods 0.000 description 8
- 238000005516 engineering process Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 239000006002 Pepper Substances 0.000 description 2
- 230000002146 bilateral effect Effects 0.000 description 2
- 235000013399 edible fruits Nutrition 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000033228 biological regulation Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000007621 cluster analysis Methods 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000009792 diffusion process Methods 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 230000001771 impaired effect Effects 0.000 description 1
- 238000003064 k means clustering Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004321 preservation Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 230000001629 suppression Effects 0.000 description 1
- 230000009466 transformation 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of non-local mean image de-noising method based on unsupervised learning, include the following steps：Image edged；Image block divides；Calculate similitude；Required pixel is estimated using weights, the value of whole image can be estimated by traveling through all search windows, finally reach the effect of denoising.The advantage of the invention is that：The method that unsupervised learning clusters is applied to image denoising first, operational efficiency of the non-local mean in image denoising is increased substantially, effectively raises the effect of denoising.The method of cluster flocks together homogeneity pixel, and calculating similarity weight is carried out using cluster centre, reduces the number for the pixel for participating in calculating, and lifts processing speed.Meanwhile only for selection cluster centre to homogeneous data denoising, influencing each other between reduction is inhomogeneous, enables edge preferably to preserve.Therefore this algorithm, which has, calculates the time few, high treating effect, the characteristics of preserving more original image informations.
Description
Technical field
The present invention relates to Image Denoising Technology field, more particularly to a kind of non-local mean image based on unsupervised learning
Denoising method.
Background technology
Among image procossing, image denoising is very important preconditioning technique, it is other technologies in image procossing
Basis and premise, image denoising be to extract original image from impaired image, carried for other stages of image procossing
For clearly, completely, be as closely as possible to the image information of original image.So efficient, image denoising side of high quality
Method, the status in image procossing are particularly significant.Therefore, image denoising is all very popular one in image procossing all the time
A research topic.Up to the present, many very effective methods are suggested applied to image denoising.Wherein, it is non local
Mean Filtering Algorithm is exactly one of more representational method in these algorithms, be widely used in remote sensing images denoising,
The denoising of scene image or video and the denoising of medical image, and have good denoising effect.It is by making full use of
The redundancy of image in itself, the correlation of pixel in image is searched in global scope, the computational accuracy of weights is improved, makes
The estimate of pixel more close to the original value of image, achievees the purpose that to remove noise, this method can be effective in addition
Remove the detailed information that noise preserves more original images.
Non-local mean filtering scans in global scope, and original image is split as size identical image block, point
Similitude that Ji Suan be between image block, equalization similarity weight is so as to estimating the value of corresponding pixel, the method
Solves the problem of traditional local denoising method, noise removal capability is weak or loses excessive raw information well.It is non local equal
Value filtering algorithm carries out calculating similitude in global scope, preserves the detailed information of more images, from the meter of similarity weight
We can see that this method brings the expense more calculated at the same time in calculation mode.Non-local mean filtering algorithm can divide
For three steps：The first step, selects a certain size search window, is split image according to the size of search window；Second
Image, in same search window, is divided into the image block of formed objects, calculates the image block where search window central point by step
With the similarity weight of all image blocks in same search window, the calculating of similarity weight is equalized as the pixel in block
The Euclidean distance of point obtains；3rd step, the pixel value of search window central point is estimated by similarity weight.
Unsupervised learning is a kind of instrument for being widely used in deep learning and machine learning, while it is to carry out data
One of main instrument excavated, unsupervised method is also used among image procossing.Cluster analysis belongs to unsupervised learning
One kind of method, it is that same class data are gathered one as far as possible using the similitude of data under conditions of no label
Inside class, while it can also allow the difference between different classes big as far as possible.Used herein is k-means algorithms,
It is the highest algorithm of accuracy rate that Adam Coates et al. demonstrate K-means clustering algorithms in their paper, and is not required to
Want hyper parameter (hyper-parameter), thus select herein be cluster in K-means methods calculated.Figure
The noise of picture randomly generates, especially among actual application, due to the difference between equipment and external condition not
It is different, even different Image Acquisition person with the noise produced, the noise included in the image being collected into
Have difference, it is desirable to which it is extremely difficult to find a kind of general treating method, and the unsupervised learning proposed in this paper that is based on (gathers
Class) image go the method made, exactly making full use of collected image, possessed characteristic is clustered in itself, is found out
Maximum difference between general character between image interior pixels point, and each class.Above said, the process of denoising is in itself
It is exactly a method using equalization, restored image is extracted from noise image, the effect of denoising is reached with this, so
The method of the k-means of selection also can just realize denoising, so the method proposed in theory is feasible.
The prior art：
Image denoising is all extremely important all the time, and is the popular research topic of comparison, the side for having many denoisings
Method and model are suggested applied to remote sensing images, scene image and medical image, while achieve certain effect.It includes
Local, non local denoising method, linear, nonlinear denoising method, frequency domain, spatial domain denoising method, is also wrapped
Include wavelet transformation, anisotropy and with Anisotropic diffusion etc..
Mean filter, according to the needs of actual conditions, selects different size of image block, with the figure where pixel to be asked
As the average of block estimates pixel value, this method can be imitated according to selecting different size of image block obtain different denoisings
Fruit, using the method for equalization, can effectively remove the influence of noise on image.
Medium filtering, it is similar to mean filter, and select a certain size image block.Unlike, medium filtering is
With going to estimate required pixel value in image block, this method can be good at excluding shadow of the salt-pepper noise to image
Ring.
Neighborhood Filtering, it is contemplated that more or less there are correlation, Neighborhood Filtering between adjacent image block to utilize this phase
Closing property estimates the value of pixel.In same image block, Neighborhood Filtering is according to pixel to be asked with being used for what is estimated
The distance relation of pixel, distributes a weights, distance pixel to be asked is more remote to each pixel for being used to estimate
Weights are smaller.This method considers the position relationship between pixel, is that denoising effect has arrived lifting.
Gaussian filtering, similar to Neighborhood Filtering, gaussian filtering is also to distribute a weights, power to the pixel for estimation
The method of salary distribution of value is different from Neighborhood Filtering, and gaussian filtering carries out the distribution of weights using the degree of similarity of pixel value in itself,
The same effect that can reach denoising.
Bilateral filtering, it is contemplated that there is mutual pass in the position between pixel in itself in same image block internal pixel values
Connection property, bilateral filtering is then that Neighborhood Filtering is combined with two methods of gaussian filtering to reach denoising purpose.
Non local filtering algorithm is not limited to consider the relation between the pixel inside single image block, but in the overall situation
In the range of, image is analyzed, the relation between search pixel point, calculate point all in image and carrying out some pixel
The estimation of value is done contribution.Neighborhood where all pixels and its is formed an image block by this method, all
Point is all inside an equal-sized image block, image block and other all pixels where the pixel to be estimated of calculating
The Euclidean distance of image block where point, for representing the similitude weights of two pixels.Non local method, in global model
Interior search is enclosed, saves the detailed information of image well, there is good denoising effect.
The prior art also has made some improvements in non local method, not every when calculating similitude weights
Point all carries out the calculating of similitude weights, they utilize SVD methods, and primary screening work is carried out to the pixel for participating in estimation,
The less pixel of correlation is directly excluded to be not involved in calculating, this method reduces influence mutual inside pixel,
Under higher noise intensity, the effect of denoising is obviously improved this method, achieves the effect become reconciled.
The prior art also adds a method for being similar to pretreatment before non local filtering, is removed using medium filtering
The salt-pepper noise in image has been removed, has excluded the larger point of influence of noise, then noise image has been carried out again using wiener filtering
Once filtering is that image is smoothened.Most handled using non-local mean filtering, this method has the performance of denoising
Further lifting.
Also using a kind of non local method abstract at random, the denoising applied to image.This method is similar in calculating
Property weights process among, using the method for random sampling, greatly reduce and participate in the pixel that similarity weight calculates
Number, good effect is achieved in the efficiency of lifting denoising.
The method of image denoising has very much, also has much for the innovatory algorithm of a certain denoising method, listed above
It is more commonly used method and method and its improved method relatively good on denoising effect, every kind of method has the advantages of uniqueness
With corresponding deficiency.
The shortcomings of the prior art：
1. simple method denoising effect is bad
These methods such as mean filter, principle is simple, it is convenient to realize, but there is denoising performance it is bad the problem of.Filtering
Intensity is excessive, then the raw information of many images can be lost while noise is removed；Filtering strength is small, then noise less than very
Good elimination.Need adjusting filtering strength cumbersome during denoising, nevertheless, sometimes still reaching to less than desired denoising effect
Fruit.
2. the time that preferable denoising effect is spent is more
It is known that with the lifting of technological means, the resolution ratio of the image acquired in us is higher and higher, data volume
It is very huge, carry out image procossing and require a great deal of time, non local method denoising effect is preferable, but needs to travel through
Global all pixels, are estimated just to require a great deal of time, to a pixel sometimes time during telephone expenses very
What hardly possible received.
3. improved method is not ideal enough
Above we refer to, and also have much for the improved method of non-local mean, but all be pair but on the one hand
Improvement, have lost the efficiency of calculating while the denoising effect of boosting algorithm, the efficiency of calculating is improved but goes
The performance made an uproar has been got back suppression, and two aspects cannot be occasionally complementary to one another well, be a comparison stubborn problem.
The content of the invention
A kind of the defects of present invention is directed to the prior art, there is provided non-local mean image denoising based on unsupervised learning
Method, can effectively solve the problem that the above-mentioned problems of the prior art.
In order to realize above goal of the invention, the technical solution that the present invention takes is as follows：
A kind of non-local mean image de-noising method based on unsupervised learning, includes the following steps：
Step 1：Image edged, before being operated to image, a frame, the size of frame are added to original image
For the selected radius for being used to calculate the image block of similitude weights.
Step 2：Image block divides, and the division of image block is divided into：Search window and neighborhood window, two kinds of image
Block, search window provide the quantity of the image block for estimating, the image in neighborhood window is used to calculate similitude weights.
Step 3：Similitude is calculated, calculates similitude weights for the estimation to image slices vegetarian refreshments；Similitude is calculated to be divided into
Two stages：First, clustered using k-means, second, calculating the Euclidean distance of pixel value.
Step 4：Required pixel is estimated using weights, is searched in same rope in window, each neighborhood window pair
There are a similitude weights in the neighborhood window where central point, represented with the similitude between neighborhood window in neighborhood window
The similitude of heart point, computational methods are：
Each search window is estimated that the value of the value of a pixel, i.e. search window central point, and traversal is all
Search window can estimate the value of whole image, finally reach the effect of denoising.
Further, the division of search window is the size of search window takes in this project experience in the step 2
Value 2, divides an image into the image block that radius is 2, the center of each image block is the pixel to be estimated, original image
Each pixel in a search window.
Further, the division of neighborhood window is by all pixels point in ready-portioned search window in the step 2
Distribute in a neighborhood window, the size empirical value of neighborhood window takes 3, in search window all points have one it is corresponding
Neighborhood window, and all neighborhood window sizes are equal.
Further, k-means clusters are to carry out all neighborhood windows using k-means methods in the step 3
Cluster, obtains a certain number of cluster centres, and same uses empirical value 5；The maximum of cluster centre is in neighborhood window
The number of pixel.
Further, it is that each neighborhood window has 5 after completing cluster operation that similitude weights are calculated in the step 3
A cluster centre, neighborhood window bag of the neighborhood window where with other points in same search window where calculating central point
The Euclidean distance of itself is included, computational methods are：
Compared with prior art the advantage of the invention is that：The method that unsupervised learning clusters is gone applied to image first
Make an uproar, increase substantially operational efficiency of the non-local mean in image denoising, effectively raise the effect of denoising.The side of cluster
Method flocks together homogeneity pixel, and calculating similarity weight is carried out using cluster centre, reduces the pixel for participating in calculating
The number of point, lifts processing speed.Meanwhile cluster centre is only selected to homogeneous data denoising, the phase between reduction is inhomogeneous
Mutually influence, edge is occasionally preferably preserved.Therefore this algorithm is few with the calculating time, high treating effect, preserves more former
The characteristics of beginning image information.
Brief description of the drawings
Fig. 1 is the flow chart of the embodiment of the present invention；
Fig. 2 is original image schematic diagram；
Fig. 3 is the schematic diagram that the embodiment of the present invention handles image a；
Fig. 4 is the schematic diagram that the embodiment of the present invention handles image b；
Fig. 5 is the schematic diagram that the embodiment of the present invention handles image c；
Fig. 6 is the schematic diagram that the embodiment of the present invention handles image d；
Fig. 7 is the schematic diagram that the embodiment of the present invention handles image e；
Fig. 8 is the schematic diagram that the embodiment of the present invention handles image f；
Fig. 9 is the schematic diagram that the embodiment of the present invention handles image g.
Embodiment
For the objects, technical solutions and advantages of the present invention are more clearly understood, develop simultaneously embodiment referring to the drawings, right
The present invention is described in further details.
As shown in Figure 1, a kind of non-local mean image de-noising method based on unsupervised learning, includes the following steps：
Step 1：Image edged, in order to facilitate handling image below, before being operated to image, to original
Beginning image adds a frame, and the size of frame is the selected radius for being used to calculate the image block of similitude weights.So
The operation for carrying out next step again is not have to distinguishing the boundary member of image and center section, is conducive to programming and realizes.
Here it is to be realized because edged can simplify programming to one side, conveniently using edged as a single part
The pixel of boundary member is handled, in addition edged is bigger to the influential effect of image denoising, selects rational edged side
Formula advantageously ensures that the effect of denoising.
Fig. 2 represents original image, and each of which grid represents a pixel, size 15x15.Fig. 3 represents edged
Image afterwards, the right and left is different from figure, and the neighborhood windows radius of selection is 3, and the size of all edgeds is also 3.
The right and left is inconsistent as seen from the figure, represents the mode of two kinds of different edgeds respectively.The left side is directly by outermost picture
Vegetarian refreshments directly replicates, and the right is an image copying, and two kinds of edged methods are more commonly used and closer to the effect for being really, so
List herein.The mode such as 0 is directly mended in addition, also having.The side of left and right, the mode phase on upper and lower benefit side are simply added in Fig. 3
Seemingly, according to looking for same process to be handled.Illustrate a little in same image, it is preferred to use an edged mode.
Step 2：Image block divides
The division of image block is divided into：Search window and neighborhood window, two kinds of image block, search window regulation are used for
The quantity of the image block of estimation, the image in neighborhood window are used to calculate similitude weights.
Step 21：The division of search window, the empirical value 2 that the size of rustling sound window takes in this project.Image is divided
The image block for being 2 for radius, the center of each image block are the pixels to be estimated, each pixel of original image
In a search window.
Image after edged is expanded to 21x21 by original 15x15, and each pixel is allocated one centered on the point
Search window in, radius 2.On border and the place of angle point without considering the pixel on side is mended, as shown in Figure 4.
Step 22：The division of neighborhood window, distributes all pixels point in ready-portioned search window to a neighborhood window
In mouthful, the size empirical value of neighborhood window takes 3.All points have a corresponding neighborhood window, and institute in search window
Some neighborhood window sizes are equal.
As shown in figure 5, each pixel is divided into a neighborhood window, and the size of all neighborhood windows is all
It is equal, for estimating the similarity weight of central point.
Step 3：Similitude is calculated, calculates similitude weights for the estimation to image slices vegetarian refreshments.The calculating of weights is this
The core procedure of method, while be also one step of key that this method is different from other methods.It is divided into two stages：First, use k-
Means is clustered, second, calculating the Euclidean distance of pixel value.
Step 31：As shown in fig. 6, k-means is clustered, all neighborhood windows are clustered using k-means methods,
A certain number of cluster centres are obtained, same uses empirical value 5.The maximum of cluster centre is pixel in neighborhood window
Number.
Each neighborhood window is clustered using k-means, obtains 5 cluster centres, the center of cluster can be with feed side
The distance between boundary, can highlight border.As shown in fig. 7, the image on the left side draws image into crossing with this cluster operation
It is divided into three classes, the border between class and class is apparent to be represented, then carry out equalization when can still preserve more
More detailed information, since such processing step causes this method to have preferable denoising performance.K-means clusters are fast in itself
Quickly, the almost operation to program does not have an impact degree.Cluster centre replaces pixel to calculate similarity weight, reduces computing
Amount.
Step 32：Similitude weights are calculated, each neighborhood window has 5 cluster centres after completing cluster operation,
Neighborhood window where central point is calculated in same search window includes the Euclidean of itself with the neighborhood window where other points
Distance.Computational methods can be expressed as：
As shown in figure 8, the method for non local filtering, considers the correlation between pixel in the overall situation, consider each
The contribution that pixel makes the pixel value to be estimated, makes the pixel value that estimates more accurate, can be good at preservation figure
As original detailed information.
Step 4：Required pixel is estimated using weights
As shown in figure 9, being searched in same rope in window, each neighborhood window has for the neighborhood window where central point
One similitude weights, the similitude of neighborhood window center point is represented with the similitude between neighborhood window.Similarity weight is
During to scan for the estimation of window center point, contribution which is made.Computational methods can be with side for：
Each search window is estimated that the value of the value of a pixel, i.e. search window central point, and traversal is all
Search window can estimate the value of whole image, finally reach the effect of denoising.
Those of ordinary skill in the art will understand that the embodiments described herein, which is to help reader, understands this hair
Bright implementation, it should be understood that protection scope of the present invention is not limited to such special statement and embodiment.Ability
The those of ordinary skill in domain can according to the present invention these disclosed technical inspirations make it is various do not depart from essence of the invention its
Its various specific deformations and combination, these deformations and combination are still within the scope of the present invention.
Claims (5)
1. a kind of non-local mean image de-noising method based on unsupervised learning, it is characterised in that include the following steps：
Step 1：Image edged, before being operated to image, adds a frame, the size of frame is institute to original image
That chooses is used to calculate the radius of the image block of similitude weights；
Step 2：Image block divides, and the division of image block is divided into：Search window and neighborhood window, two kinds of image block, is searched
Rope window provides the quantity of the image block for estimating, the image in neighborhood window is used to calculate similitude weights；
Step 3：Similitude is calculated, calculates similitude weights for the estimation to image slices vegetarian refreshments；Calculate similitude and be divided into two
Stage：First, clustered using k-means, second, calculating the Euclidean distance of pixel value；
Step 4：Required pixel is estimated using weights, searched in same rope in window, each neighborhood window is in
Neighborhood window where heart point has a similitude weights, and neighborhood window center point is represented with the similitude between neighborhood window
Similitude, computational methods are：
<mrow>
<mi>I</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>&Element;</mo>
<mi>N</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</munder>
<mi>w</mi>
<mrow>
<mo>(</mo>
<msub>
<mover>
<mi>x</mi>
<mo>&OverBar;</mo>
</mover>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mover>
<mi>x</mi>
<mo>&OverBar;</mo>
</mover>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
<msub>
<mi>I</mi>
<mi>n</mi>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>;</mo>
</mrow>
Each search window is estimated that the value of the value of a pixel, i.e. search window central point, travels through all search
Rope window can estimate the value of whole image, finally reach the effect of denoising.
2. a kind of non-local mean image de-noising method based on unsupervised learning according to claim 1, its feature exist
In：The division of search window is the empirical value 2 that the size of search window takes in this project in the step 2, and image is divided
The image block for being 2 for radius, the center of each image block are the pixels to be estimated, each pixel of original image
In a search window.
3. a kind of non-local mean image de-noising method based on unsupervised learning according to claim 1, its feature exist
In：The division of neighborhood window is to distribute all pixels point in ready-portioned search window to a neighborhood window in the step 2
In mouthful, the size empirical value of neighborhood window takes 3, and all points have a corresponding neighborhood window, and institute in search window
Some neighborhood window sizes are equal.
4. a kind of non-local mean image de-noising method based on unsupervised learning according to claim 1, its feature exist
In：K-means clusters are to be clustered all neighborhood windows using k-means methods in the step 3, obtain a fixed number
The cluster centre of amount, same uses empirical value 5；The maximum of cluster centre is the number of pixel in neighborhood window.
5. a kind of non-local mean image de-noising method based on unsupervised learning according to claim 4, its feature exist
In：It is that each neighborhood window has 5 cluster centres after completing cluster operation that similitude weights are calculated in the step 3,
Neighborhood window where central point is calculated in same search window includes the Euclidean of itself with the neighborhood window where other points
Distance, computational methods are：
<mrow>
<mi>W</mi>
<mrow>
<mo>(</mo>
<msub>
<mover>
<mi>x</mi>
<mo>&OverBar;</mo>
</mover>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mover>
<mi>x</mi>
<mo>&OverBar;</mo>
</mover>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mrow>
<mi>Z</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<msup>
<mrow>
<mo>(</mo>
<mi>E</mi>
<mo>(</mo>
<mi>x</mi>
<mi>i</mi>
<mo>)</mo>
<mo>-</mo>
<mi>E</mi>
<mo>(</mo>
<mi>x</mi>
<mi>j</mi>
<mo>)</mo>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
<msup>
<mi>h</mi>
<mn>2</mn>
</msup>
</mfrac>
<mo>)</mo>
</mrow>
<mo>.</mo>
</mrow>
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711312994.0A CN107909558A (en) | 2017-12-12 | 2017-12-12 | A kind of non-local mean image de-noising method based on unsupervised learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711312994.0A CN107909558A (en) | 2017-12-12 | 2017-12-12 | A kind of non-local mean image de-noising method based on unsupervised learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107909558A true CN107909558A (en) | 2018-04-13 |
Family
ID=61865204
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711312994.0A Pending CN107909558A (en) | 2017-12-12 | 2017-12-12 | A kind of non-local mean image de-noising method based on unsupervised learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107909558A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111553860A (en) * | 2020-04-29 | 2020-08-18 | 北京理工大学 | Deep learning non-neighborhood averaging processing method and system for water color remote sensing image |
| CN111667424A (en) * | 2020-05-28 | 2020-09-15 | 武汉大学 | Unsupervised real image denoising method |
| CN112330566A (en) * | 2020-11-12 | 2021-02-05 | 武汉工程大学 | Image denoising method and device and computer storage medium |
| CN112419198A (en) * | 2020-11-27 | 2021-02-26 | 中国矿业大学 | Non-local mean weighting method for SAR interferogram filtering |
| CN113034387A (en) * | 2021-03-05 | 2021-06-25 | 成都国科微电子有限公司 | Image denoising method, device, equipment and medium |
| CN116703920A (en) * | 2023-08-07 | 2023-09-05 | 济宁市华祥石墨制品有限公司 | Graphite thermal insulation barrel crack detection method based on image processing |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160098820A1 (en) * | 2014-10-03 | 2016-04-07 | Raghu Kopalle | System for robust denoising of images |
| CN106600567A (en) * | 2017-01-03 | 2017-04-26 | 长沙全度影像科技有限公司 | Adaptive non-local mean denoising method |
| US9659352B2 (en) * | 2012-10-31 | 2017-05-23 | Google Inc. | Image denoising system and method |
| CN106875355A (en) * | 2017-01-19 | 2017-06-20 | 合肥工业大学 | The non local noise-reduction method of polarization SAR of homogeneity pixel selection is clustered based on bicharacteristic |
-
2017
- 2017-12-12 CN CN201711312994.0A patent/CN107909558A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9659352B2 (en) * | 2012-10-31 | 2017-05-23 | Google Inc. | Image denoising system and method |
| US20160098820A1 (en) * | 2014-10-03 | 2016-04-07 | Raghu Kopalle | System for robust denoising of images |
| CN106600567A (en) * | 2017-01-03 | 2017-04-26 | 长沙全度影像科技有限公司 | Adaptive non-local mean denoising method |
| CN106875355A (en) * | 2017-01-19 | 2017-06-20 | 合肥工业大学 | The non local noise-reduction method of polarization SAR of homogeneity pixel selection is clustered based on bicharacteristic |
Non-Patent Citations (2)
| Title |
|---|
| ASIFULLAH KHAN ET AL.: "Image de-noise using noise ratio extimation，k-means clustering and non-local means-based estimation", 《COMPUTERS AND ELECTRICAL ENGINEERING》 * |
| 王林: "基于非局部均值的图像去噪方法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111553860A (en) * | 2020-04-29 | 2020-08-18 | 北京理工大学 | Deep learning non-neighborhood averaging processing method and system for water color remote sensing image |
| CN111553860B (en) * | 2020-04-29 | 2023-04-28 | 北京理工大学 | Deep learning non-neighborhood averaging processing method and system for water color remote sensing image |
| CN111667424A (en) * | 2020-05-28 | 2020-09-15 | 武汉大学 | Unsupervised real image denoising method |
| CN111667424B (en) * | 2020-05-28 | 2022-04-01 | 武汉大学 | Unsupervised real image denoising method |
| CN112330566A (en) * | 2020-11-12 | 2021-02-05 | 武汉工程大学 | Image denoising method and device and computer storage medium |
| CN112330566B (en) * | 2020-11-12 | 2023-11-14 | 武汉工程大学 | Image denoising method and device and computer storage medium |
| CN112419198A (en) * | 2020-11-27 | 2021-02-26 | 中国矿业大学 | Non-local mean weighting method for SAR interferogram filtering |
| CN112419198B (en) * | 2020-11-27 | 2024-02-02 | 中国矿业大学 | Non-local mean weighting method for SAR interferogram filtering |
| CN113034387A (en) * | 2021-03-05 | 2021-06-25 | 成都国科微电子有限公司 | Image denoising method, device, equipment and medium |
| CN113034387B (en) * | 2021-03-05 | 2023-07-14 | 成都国科微电子有限公司 | Image denoising method, device, equipment and medium |
| CN116703920A (en) * | 2023-08-07 | 2023-09-05 | 济宁市华祥石墨制品有限公司 | Graphite thermal insulation barrel crack detection method based on image processing |
| CN116703920B (en) * | 2023-08-07 | 2023-10-20 | 济宁市华祥石墨制品有限公司 | Graphite thermal insulation barrel crack detection method based on image processing |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107909558A (en) | A kind of non-local mean image de-noising method based on unsupervised learning | |
| CN108876737B (en) | Image denoising method combining residual learning and structural similarity | |
| Jain et al. | A survey of edge-preserving image denoising methods | |
| Knaus et al. | Progressive image denoising | |
| CN114140353A (en) | Swin-Transformer image denoising method and system based on channel attention | |
| CN106709877B (en) | A kind of image deblurring method based on multi-parameter canonical Optimized model | |
| CN103761719B (en) | A kind of adaptive wavelet threshold denoising method based on neighborhood relevance | |
| CN101980284A (en) | Two-scale sparse representation-based color image noise reduction method | |
| CN109003233B (en) | Image denoising method based on self-adaptive weight total variation model | |
| CN107784638A (en) | A kind of Dongba ancient books image enchancing method of optimization | |
| CN103208097A (en) | Principal component analysis collaborative filtering method for image multi-direction morphological structure grouping | |
| CN109389560A (en) | A kind of adaptive weighted filter image denoising method, device and image processing equipment | |
| CN113724164B (en) | Visible light image noise removing method based on fusion reconstruction guidance filtering | |
| CN103870820A (en) | Illumination normalization method for extreme illumination face recognition | |
| CN109345609A (en) | Mural painting image denoising is carried out based on convolutional neural networks and line drawing draws the method generated | |
| CN106651789B (en) | Self-adaptive deblocking method for compressed face image | |
| CN106097257B (en) | A kind of image de-noising method and device | |
| El Hassani et al. | Efficient image denoising method based on mathematical morphology reconstruction and the Non-Local Means filter for the MRI of the head | |
| Mrázek et al. | From two-dimensional nonlinear diffusion to coupled Haar wavelet shrinkage | |
| CN104966271B (en) | Image de-noising method based on biological vision receptive field mechanism | |
| CN116385312A (en) | Low-illumination image denoising method based on phase correlation | |
| CN107230191B (en) | Non-local mean denoising optimization method based on structural similarity | |
| Mahalakshmi et al. | Adaptive wavelet packet decomposition for efficient image denoising by using neighsure shrink method | |
| CN103310424B (en) | A kind of image de-noising method based on structural similarity Yu total variation hybrid model | |
| Ouyang et al. | Research on DENOISINg of cryo-em images based on 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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180413 |