CN107301644A - Natural image non-formaldehyde finishing method based on average drifting and fuzzy clustering - Google Patents
Natural image non-formaldehyde finishing method based on average drifting and fuzzy clustering Download PDFInfo
- Publication number
- CN107301644A CN107301644A CN201710434179.5A CN201710434179A CN107301644A CN 107301644 A CN107301644 A CN 107301644A CN 201710434179 A CN201710434179 A CN 201710434179A CN 107301644 A CN107301644 A CN 107301644A
- Authority
- CN
- China
- Prior art keywords
- mrow
- msub
- pixel
- msubsup
- mfrac
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
Abstract
The present invention is a kind of natural image non-formaldehyde finishing method based on average drifting and fuzzy clustering, mainly solves the problem of prior art is low to the non-formaldehyde finishing accuracy rate of mass natural image.Its scheme is：1) input picture, is carried out smoothly to it；2) 64 iteration initial points of equality initialization in the normalization RGB color space of smooth rear image pixel；3) search is iterated to initial point, 64 convergence points are obtained；4) delete number of pixels in the higher-dimension ball centered on convergence point and be less than the convergence point for deleting threshold value；5) merge Euclidean distance and be less than the convergence point for merging threshold value, determine density peaks and density peaks number, the degree of membership of pixel and the smooth degree of membership of pixel are calculated successively；6) to the smooth degree of membership de-fuzzy of pixel, it is that each pixel adds class label, exports segmentation figure picture.The present invention need not set control parameter, the segmentation classification number of image can be automatically determined, available for the non-formaldehyde finishing to mass natural image.
Description
Technical field
Technical field of image processing of the present invention, it is specifically a kind of to natural image non-formaldehyde finishing method, available for video mesh
Mark in Tracking Recognition and CBIR.
Background technology
In recent years, with the fast development of science and technology and computer internet technology, digital picture is in all trades and professions
What is used is more and more extensive.In large nuber of images, how to quickly recognize piece image, always as computer vision with
The topic of pattern-recognition hot discussion.Since in the early 1990s, CBIR technology is suggested, it
The always study hotspot of researcher, it is mainly by extracting texture, color, the shape of target and their space of image
The features such as positional information, calculate the similarity distance of the image in be retrieved image and data set, come realize the identification of image with
Retrieval.By the research and development of nearly 20 years, in application aspect also comparative maturity, as google, Baidu, Bing etc. receive rope
Engine company all develops the one's own picture search product based on content.Such as：Google Similar Images,
Figure etc. is known by Baidu.
There is three kinds of big datas, the i.e. text of static Web page, voice, image on the big data epoch, network media and regard
Frequently, the proportion of wherein video and image is maximum.These images or frame of video are carried out quick and precisely, unsupervised Ground Split, can produce
Raw huge economic benefit.Such as, image is split for image retrieval, image and video editing, based on image and video content
Advertisement putting.In the artificial intelligence epoch, computer vision is the important component of artificial intelligence, and image segmentation is especially certainly
Dynamic cutting techniques are the key technologies of machine vision, in terms of image recognition, target following, scene analysis identification, image point
Cut the effect for all playing key.
The segmentation classification number for automatically determining image is always the focus and difficult point of academia and industrial quarters research, is not had also at present
There is highly effective algorithm to be accurately determined the classification number of every piece image.It is relatively good for natural image simple in construction
Method can accurately determine classification number, but for the more complicated image of structure, can typically there is error.Existing natural image
Non-formaldehyde finishing method is unstable, easily causes over-segmentation or the less divided of image.Because natural image scene is more complicated, mesh
The preceding classification number that natural image is also accurately determined without relatively good method.
The content of the invention
It is a kind of based on average drifting and fuzzy clustering it is an object of the invention to for above-mentioned the deficiencies in the prior art, propose
Natural image non-formaldehyde finishing method, to accurately determine the classification number of image, improve segmentation precision.
The technical proposal of the invention is realized in this way：
One, technical principles：
The pixel of each target area in natural image spatially has certain Density Distribution shape in RGB feature
Formula, and the pixel of each classification has a density maximum point in RGB feature spatial distribution, is exactly the master of classification on image
Color.The density maximum point of image pixel distribution is referred to as density peaks, and it is actually the cluster of pixel in cut zone
Center, have found density peaks, determine that the cluster centre of cut zone, while just can be true according to the number of density peaks
Determine the cut zone number of image.According to this thinking, the present invention searches for natural image in RGB color using mean shift algorithm
The density peaks in space, the segmentation classification number and cluster centre of image can be determined simultaneously, then calculated with improved fuzzy clustering
Method realizes the segmentation of natural image, and the inventive method can realize the automatic segmentation of natural image.
Two, technical schemes
According to above-mentioned principle, it is as follows that technical scheme includes for of the invention realizing：
(1) natural image I to be split is readt, and to natural image ItThe rgb value of all pixels divided by 255 it is normalized
Processing so that the scope of each pixel RGB values is between [0,1], and wherein t=1,2 ..., n, n represents amount of images to be split；
(2) natural image after normalization is carried out smoothly, obtaining smoothed image I 't；
(3) 64 search initial points of equality initialization are initial clustering in the RGB color space of smooth rear natural image
Center, corresponding initial cluster center number c=64, the initial point set expression that these starting points are constituted is V={ v1,v2,…,
vp,…,v64, wherein, vpRepresent p-th of starting point, p=1,2 ..., 64；
(4) mean shift iterations formula is used, search is iterated using 64 initial points as starting point, obtains restraining point set
V '={ v '1,v′2,…,v′k,…,v′64, wherein, v 'kRepresent k-th of convergence point；
(5) in RGB color space, convergence point v ' is setkSubscript initial value k=1, set threshold value M=100, delete
The smaller convergence point of surrounding pixel point distribution density：
(5a) is calculated with convergence point v 'kThe pixel number n included for the centre of sphere, the higher-dimension ball by radius of hk, compare nkWith
M size：If nk＜ M, then from the convergence middle deletion convergence point v ' of point set V 'k, and make c=c-1, otherwise, do not delete v 'k；
(5b) makes k=k+1, judges whether k≤64 set up：If so, return (5a), otherwise into step (6)；
(6) convergence point set V '={ v ' is merged1,v′2,…,v′p,…,v′q,…,v′cIn between any two convergence point
Euclidean distance is less than threshold value h convergence point, obtains cluster centre set V ", wherein, v 'pRepresent p-th of convergence point, v 'qRepresent the
Q convergence point, p ≠ q；
(7) the cluster centre set V " calculated according to step (6), the natural image I ' after calculating smoothlytPixel
Point subordinated-degree matrix U, the row k of matrix U, the i-th column element ukiCalculation formula it is as follows：；
(8) subordinated-degree matrix U is carried out smooth, obtains the smooth subordinated-degree matrix U ' of pixel, matrix U ' row k, i-th
Column element u 'kiCalculation formula it is as follows：
Wherein, u 'ki∈ [0,1], i=1,2 ..., N, N represent smoothed image I 'tThe sum of middle pixel, k=1,2 ..., c,
C represents the sum of cluster centre, NiRepresent the neighborhood territory pixel set in the smooth window centered on ith pixel, sjAnd siTable
Show the space coordinate of pixel in smooth window, xiRepresent the value of smooth window center pixel, xjThe value of neighborhood territory pixel is represented, is set
The size of smooth window is 5 × 5, σsRepresent that the space nucleus band of smooth window is wide, in whole smoothing process, σsSize be solid
It is fixed constant, σ is sets=1, σirRepresent that the corresponding adaptive codomain nucleus band of ith pixel is wide, NRRepresent neighborhood in neighborhood window
The number of pixel；
(9) using smooth degree of membership u ' of the maximum membership degree rule to each pixelkiDe-fuzzy is carried out, image is obtained
ItThe class label L of middle ith pixeli：
(10) repeat step (9), to image ItIn all pixels point carry out de-fuzzy, obtain segmentation figure picture, split
The pixel of each cut zone has identical class label in image.
The present invention has advantages below compared with prior art：
1. the present invention is smoothed to image, it can effectively suppress the influence of noise and outlier to segmentation result.
2. the density peaks for the method search pixel that the present invention is risen using local density's gradient, can not only find pixel
Distribution pattern, and can determine that out the cluster centre and segmentation classification number of global optimum.
3. the present invention is smooth by being carried out to pixel subordinated-degree matrix, the suppression energy to noise and outlier had both been improved
Power, improves the uniformity inside cut zone again.
Brief description of the drawings
Fig. 1 be the present invention realize general flow chart；
Fig. 2 is the distribution form figure of the pixel of the invention in implementation process and mean shift iterations point in rgb space；
Fig. 3 is comparison diagram of the segmentation result with artificial segmentation result for the natural image for being A to numbering with the inventive method.
Fig. 4 is comparison diagram of the segmentation result with artificial segmentation result for the natural image for being B to numbering with the inventive method.
Fig. 5 is comparison diagram of the segmentation result with artificial segmentation result for the natural image for being C to numbering with the inventive method.
Fig. 6 is comparison diagram of the segmentation result with artificial segmentation result for the natural image for being D to numbering with the inventive method.
Embodiment
Reference picture 1, step is as follows for of the invention realizing：
Step one：Segmentation figure picture is treated to be normalized.
Input the natural image I as shown in Fig. 2 (a)t, i=1,2 ..., n, n represents that view data concentrates image to be split
Quantity；To image IiThe rgb value of pixel is normalized, and makes the value of each color channel between [0,1] scope, figure
2 (a) image slices vegetarian refreshments is distributed as shown in Fig. 2 (b) in normalization RGB color space；
Step 2：Using following smoothing formula to image ItCarry out smooth：
Wherein,+ 1 iterative value of kth of the center pixel of i-th of sliding window is represented, sliding window iteration terminates bar
Part isNiRepresent the set of all pixels in sliding window, siRepresent the sky of sliding window central point pixel
Between coordinate, sjRepresent the neighborhood territory pixel space coordinate in sliding window, σsRepresentation space constraint nucleus band is wide, σsIn whole smooth filter
It is changeless in wave process, σ is sets=1, σkRepresent that codomain nucleus band of the sliding window in kth time iteration is wide, σkWith
Iterative process is what is constantly adjusted, and its calculation formula is as follows：
Wherein, NRRepresent the sum of pixel in sliding window, NiThe set of all pixels point in i-th of sliding window is represented,
xjThe value of the pixel in sliding window is represented,The kth time iterative value of i-th of sliding window center pixel is represented, filter is set
Ripple sliding window size is 5 × 5, sets sliding window iteration ends threshold epsilon1=0.001.Image slices vegetarian refreshments is distributed such as after smooth
Fig. 2 (c)；
Step 3：The equality initialization mean shift iterations starting point in RGB color space.
64 density peaks search initial points are initialized in unit color cube space uniform, each initial point is one
The center of small cubes, unit cube is by 64 small cubes even partitions, and the volume of each small cubes is 1/64, small vertical
The radius of cube circumsphere isH is the iterative search mistake in whole color space
The homogeneous nucleus bandwidth used in journey, initial point set expression is：V={ v1,v2,…,vp,…,v64}；
Because the space of unit color cube can be completely covered in the circumsphere of 64 small cubes, this initialization is utilized
Mode ensure that all pixels in color cube are all searched in first time iteration, and 64 mean shift iterations are initial
Shown in the distribution such as Fig. 2 (d) of point in normalization RGB color space；
Step 4：In the normalization RGB color space of smoothed image, mean shift iterations formula iterative search picture is used
Vegetarian refreshments density peaks.
Iterative formula of the average drifting in color space is as follows：
Wherein,Represent m-th of iterative search point in kth time iterative value, m=1,2 ..., 64, Sh(yk) represent color in RGB
In the colour space withCentered on, the pixel point set that h is surrounded by the higher-dimension ball of radius, nkRepresent set Sh(yk) interior pixel
Number, Sh(yk) representation formula it is as follows：
Wherein, x represents set Sh(yk) in pixel, set mean shift iterations terminate threshold epsilon2=0.001, m-th
The convergent condition of iteration point is：Convergence point set expression is：V '={ v '1,v′2,…,v′64}；
Using 64 initial points as iterative search starting point, searched for using mean shift iterations formula in whole color space
Density peaks, mean shift vectors can be moved gradually to the big direction of pixel distribution density, until converging to local density's peak value
Point, due in mean shift iterations search procedure, being independent of each other between 64 mean shift iterations central points, and asynchronously receive
Hold back, and in RGB color space, be distributed around the mean shift iterations initial search point of part without pixel, therefore these iteration are searched
Rope starting point is restrained after the first iteration, and 64 convergence points are distributed as shown in Fig. 2 (e) in normalization RGB color space.
Step 5：In 64 convergence points, the convergence point for meeting threshold condition is deleted.
Convergence point sum c=64 is set, deletion threshold value M=100 is set, and by itself and set Sh(yk) interior pixel number
Mesh nkIt is compared：If nk＜ M, then delete convergence point from convergence point setAnd c=c-1 is made, otherwise, convergence is not deleted
PointM=1,2 ..., 64；
The convergence point yk+1, it is that the direction iterative search increased along density gradient is obtained, it is that local density is very big
It is worth point, corresponding pixel set Sh(yk) interior pixel quantity nkShould be local maxima, if nkLess than certain threshold value, say
This bright convergence point is not real local pixel density peaks, can not represent the pixel characteristic of a classification, it should delete
The less convergence point of these density ratios, the delet method can be prevented effectively from the over-segmentation of image, delete the convergence for the condition that meets
Obtained set expression is after point：V '={ v '1,v′2,…,v′p,…,v′q,…,v′c}。
Step 6：Merge Euclidean distance and be less than the convergence point for merging threshold value.
Set and merge threshold valueMerge set V '={ v '1,v′2,…,v′p,…,v′q,…,v′cIn Euclidean
Distance is less than threshold value h any two convergence point, and implementation process is as follows：
(6a) sets left convergence point v 'pSubscript initial value p=1；
(6b) sets right convergence point v 'qSubscript initial value q=p+1；
(6c) judges whether q≤c sets up：If so, (6d) then is performed, otherwise, is performed (6f)；
(6d) calculates d=| | v 'p-v′q| |, judge whether d ＜ h set up：If so, then make vq'=(v 'p+v′q)/2, c=
C-1, p=p+1, are returned (6b), otherwise, are performed (6e)；
(6e) makes q=q+1, judges whether q≤c sets up：If so, (6d) then is returned, otherwise, is performed (6f)；
(6f) makes p=p+1, judges whether p≤c-1 sets up：If so, (6b) then is returned, otherwise, terminates to calculate.
Said process finally calculates cluster centre set, is expressed as：V "={ v1,v2,…,vc, delete and merge convergence
The cluster centre obtained after point is distributed as shown in Fig. 2 (f) in normalization RGB color space；
Step 7：The cluster centre set V " obtained according to step 6, calculates the pixel core fuzzy membership of smoothed image.
(7a) calculates the corresponding wide σ of nucleus band of each cluster centrek：
Wherein, N represents the sum of image pixel, xjRepresent j-th of pixel of image, vkRepresent k-th of cluster centre；
(7b) is according to formula<9>The σ calculatedk, utilize formula<10>Gaussian kernel is calculated apart from k (xi,vk)：
Wherein, xiRepresent ith pixel in image；
(7c) utilizes formula<11>Calculate the subordinated-degree matrix U, u of image pixelkiFor row k in U, the degree of membership of the i-th row：
Step 8：Utilize formula<11>The subordinated-degree matrix U calculated, according to formula<12>Calculate the smooth person in servitude of image pixel
Category degree matrix U ' middle row k, the smooth degree of membership u ' of the i-th rowki：
Wherein, NiRepresent the set of pixel in the smooth window centered on pixel i, NRRepresent pixel in smooth window
Sum, sjAnd siRepresent the plane space coordinate of pixel in window, xiRepresent smooth window center pixel, xjRepresent neighborhood territory pixel,
ukjRepresent the degree of membership of j-th of pixel in smooth window, σs′Represent that the space nucleus band of smooth window is wide, σs′Whole smoothed
It is fixed in journey, takes σs′=1, σirRepresent that the corresponding codomain nucleus band of i-th of smooth window is wide, it is smooth good that pixel is subordinate to
Place is robustness of the enhancing to noise and outlier, and the continuity preferably inside protection image detail and cut zone.
Step 9：Using smooth degree of membership u ' of the maximum membership degree rule to each pixelkiDe-fuzzy is carried out, is obtained
The class label L of ith pixel in imagei：
Step 10：Segmentation figure picture is exported, next image to be split is inputted, repeats the above steps, until all images point
Cut and finish.
The effect of the present invention can be further illustrated by following experiment：
1. the image that emulation experiment is used：
Experiment uses Berkeley University natural image partitioned data set BSD300, and the data set has 300 width natural images, this
Invention carries out non-formaldehyde finishing to this 300 width natural image, and 4 width segmentation figure pictures are chosen in 300 width segmentation figure pictures in specification
It is shown in accompanying drawing, this four width picture number is respectively A, B, C and D.
2. the parameter setting of emulation experiment：
If the sliding window iteration convergence threshold epsilon for image smoothing1=0.001, the sliding window space wide σ of nucleus bands=1,
Sliding window size is 5 × 5, average drifting codomain bandwidthPoint deletion threshold value M=100 is restrained, convergence point merges
Threshold valueMean shift iterations convergence threshold ε2=0.001, the smooth window space smooth for subordinated-degree matrix
The wide σ of nucleus bands′=1.
3. emulation experiment environment：
CPU is that core3 3.2GHz, program running memory are that 4Gb, operating system are Windows7 systems, program operation ring
Border is OpenCV+VisualStudio2010.
4. emulation content：
Emulation 1, is that A natural images are split to numbering with the inventive method, and enter with the standard results manually split
Row contrast, segmentation result such as Fig. 3, wherein：
Fig. 3 (a) is the standard reference results cut in the enterprising pedestrian's work points of image A,
Fig. 3 (b) is the artificial binary border segmentation result to image A,
Fig. 3 (c) is direct segmentation result of the present invention to image A,
Fig. 3 (d) is binary border segmentation result of the present invention to image A；
Emulation 2, is that B natural images are split to numbering with the inventive method, and enter with the standard results manually split
Row contrast, segmentation result such as Fig. 4, wherein：
Fig. 4 (a) is the standard reference results cut in the enterprising pedestrian's work points of image B,
Fig. 4 (b) is the artificial binary border segmentation result to image B,
Fig. 4 (c) is direct segmentation result of the present invention to image B,
Fig. 4 (d) is binary border segmentation result of the present invention to image B；
Emulation 3, is that C natural images are split to numbering with the inventive method, and enter with the standard results manually split
Row contrast, segmentation result such as Fig. 5, wherein：
Fig. 5 (a) is the standard reference results cut in the enterprising pedestrian's work points of image C,
Fig. 5 (b) is the artificial binary border segmentation result to image C,
Fig. 5 (c) is direct segmentation result of the present invention to image C,
Fig. 5 (d) is binary border segmentation result of the present invention to image C；
Emulation 4, is that D natural images are split to numbering with the inventive method, and enter with the standard results manually split
Row contrast, segmentation result such as Fig. 6, wherein：
Fig. 6 (a) is the standard reference results cut in the enterprising pedestrian's work points of image D,
Fig. 6 (b) is the artificial binary border segmentation result to image D,
Fig. 6 (c) is direct segmentation result of the present invention to image D,
Fig. 6 (d) is binary border segmentation result of the present invention to image D；
The inventive method can accurately determine the segmentation classification number of natural image, the segmentation knot of the natural image of the present invention point
Fruit is much like with artificial segmentation result, illustrates that the inventive method has higher segmentation accuracy rate, segmentation result of the present invention and people
Work segmentation result is compared, and can be partitioned into the details area of image, with preferable details protective capability.
Claims (4)
1. the natural image non-formaldehyde finishing method based on average drifting and fuzzy clustering, including：
(1) to be split natural image I of the input as shown in Fig. 2 (a) is readt, and to natural image ItThe rgb value of all pixels is removed
It is normalized with 255 so that the scope of each pixel RGB values is between [0,1], and wherein t=1,2 ..., n, n is represented
Shown in the distribution such as Fig. 2 (b) of pixel in normalization RGB color space in amount of images to be split, Fig. 2 (a) images；
(2) natural image after normalization is carried out smoothly, obtaining smoothed image It', smoothed image It' pixel in normalizing
Change shown in distribution such as Fig. 2 (c) in RGB color space；
(3) 64 search initial points of equality initialization are initial cluster center in the RGB color space of smooth rear natural image,
Corresponding initial cluster center number c=64, the initial point set expression that these starting points are constituted is V={ v1,v2,…,
vp,…,v64, wherein, vpRepresent p-th of starting point, p=1,2 ..., 64, point of the initial point in normalization RGB color space
Shown in cloth such as Fig. 2 (d)；
(4) use mean shift iterations operator, search be iterated using 64 initial points as starting point, obtain restraining point set V '=
{v′1,v′2,…,v′p,…,v′q,…,v′64, wherein, v 'pRepresent p-th of convergence point, v 'qRepresent q-th of convergence point, p ≠ q,
Shown in the distribution such as Fig. 2 (e) of convergence point in normalization RGB color space；
(5) in RGB color space, convergence point v ' is setkSubscript initial value k=1, set threshold value M=100, delete surrounding picture
The smaller convergence point of vegetarian refreshments distribution density：
(5a) is calculated with convergence point v 'kThe pixel number n included for the centre of sphere, the higher-dimension ball by radius of hk, compare nkWith M's
Size：If nk＜ M, then from the convergence middle deletion convergence point v ' of point set V 'k, and make c=c-1, otherwise, do not delete v 'k；
(5b) makes k=k+1, judges whether k≤64 set up：If so, return (5a), otherwise into step (6)；
(6) convergence point set V '={ v ' is merged1,v′2,…,v′p,…,v′q,…,v′cIn Euclidean between any two convergence point
Distance is less than threshold value h convergence point, obtains cluster centre set V "={ v "1,v″2,…,v″c, wherein, v 'pRepresent p-th of receipts
Hold back a little, v 'qRepresent q-th of convergence point, p ≠ q, distribution such as Fig. 2 (f) institute of the cluster centre point in normalization RGB color space
Show；
(7) the cluster centre set V " calculated according to step (6), the natural image I ' after calculating smoothlytPixel be subordinate to
Spend matrix U, the row k of matrix U, the i-th column element ukiCalculation formula it is as follows：；
<mrow>
<msub>
<mi>u</mi>
<mrow>
<mi>k</mi>
<mi>i</mi>
</mrow>
</msub>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mrow>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>c</mi>
</munderover>
<msup>
<mrow>
<mo>(</mo>
<mfrac>
<mrow>
<mn>1</mn>
<mo>-</mo>
<mi>k</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mi>v</mi>
<mi>k</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<mn>1</mn>
<mo>-</mo>
<mi>k</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>,</mo>
<msub>
<mi>v</mi>
<mi>k</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
<mrow>
<mn>1</mn>
<mo>/</mo>
<mi>m</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msup>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mo><</mo>
<mn>1</mn>
<mo>></mo>
</mrow>
(8) subordinated-degree matrix U is carried out smooth, obtains the smooth subordinated-degree matrix U ' of pixel, matrix U ' row k, the i-th row member
Plain u 'kiCalculation formula it is as follows：
<mrow>
<msubsup>
<mi>u</mi>
<mrow>
<mi>k</mi>
<mi>i</mi>
</mrow>
<mo>&prime;</mo>
</msubsup>
<mo>=</mo>
<mfrac>
<mrow>
<munder>
<mo>&Sigma;</mo>
<mrow>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>&Element;</mo>
<msub>
<mi>N</mi>
<mi>i</mi>
</msub>
</mrow>
</munder>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>s</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<msub>
<mi>s</mi>
<mi>i</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<mrow>
<mn>2</mn>
<msubsup>
<mi>&sigma;</mi>
<mi>s</mi>
<mn>2</mn>
</msubsup>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<mrow>
<mn>2</mn>
<msubsup>
<mi>&sigma;</mi>
<mrow>
<mi>i</mi>
<mi>r</mi>
</mrow>
<mn>2</mn>
</msubsup>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
<msub>
<mi>u</mi>
<mrow>
<mi>k</mi>
<mi>j</mi>
</mrow>
</msub>
</mrow>
<mrow>
<munder>
<mo>&Sigma;</mo>
<mrow>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>&Element;</mo>
<msub>
<mi>N</mi>
<mi>i</mi>
</msub>
</mrow>
</munder>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>s</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<msub>
<mi>s</mi>
<mi>i</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<mrow>
<mn>2</mn>
<msubsup>
<mi>&sigma;</mi>
<mi>s</mi>
<mn>2</mn>
</msubsup>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<mrow>
<mn>2</mn>
<msubsup>
<mi>&sigma;</mi>
<mrow>
<mi>i</mi>
<mi>r</mi>
</mrow>
<mn>2</mn>
</msubsup>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mo><</mo>
<mn>2</mn>
<mo>></mo>
</mrow>
<mrow>
<msub>
<mi>&sigma;</mi>
<mrow>
<mi>i</mi>
<mi>r</mi>
</mrow>
</msub>
<mo>=</mo>
<msqrt>
<mrow>
<mfrac>
<mn>1</mn>
<msub>
<mi>N</mi>
<mi>R</mi>
</msub>
</mfrac>
<munder>
<mo>&Sigma;</mo>
<mrow>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>&Element;</mo>
<msub>
<mi>N</mi>
<mi>i</mi>
</msub>
</mrow>
</munder>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
</msqrt>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mo><</mo>
<mn>3</mn>
<mo>></mo>
</mrow>
Wherein, u 'ki∈ [0,1], i=1,2 ..., N, N represent smoothed image ItThe sum of ' middle pixel, k=1,2 ..., c, c tables
Show the sum of cluster centre, NiRepresent the neighborhood territory pixel set in the smooth window centered on ith pixel, sjAnd siRepresent
The space coordinate of pixel, x in smooth windowiRepresent the value of smooth window center pixel, xjThe value of neighborhood territory pixel is represented, sets flat
The size of sliding window mouthful is 5 × 5, σsRepresent that the space nucleus band of smooth window is wide, in whole smoothing process, σsSize be fixed
Constant, σ is sets=1, σirRepresent that the corresponding adaptive codomain nucleus band of ith pixel is wide, NRRepresent neighborhood picture in neighborhood window
The number of element；
(9) using smooth degree of membership u ' of the maximum membership degree rule to each pixelkiDe-fuzzy is carried out, image I is obtainedtIn
The class label L of ith pixeli：
<mrow>
<msub>
<mi>L</mi>
<mi>i</mi>
</msub>
<mo>=</mo>
<munder>
<mi>arg</mi>
<mi>k</mi>
</munder>
<mrow>
<mo>(</mo>
<mi>m</mi>
<mi>a</mi>
<mi>x</mi>
<mo>(</mo>
<msubsup>
<mi>u</mi>
<mrow>
<mi>k</mi>
<mi>i</mi>
</mrow>
<mo>&prime;</mo>
</msubsup>
<mo>)</mo>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mo><</mo>
<mn>4</mn>
<mo>></mo>
</mrow>
(10) repeat step (9), to image ItIn all pixels point carry out de-fuzzy, obtain in segmentation figure picture, segmentation figure picture
The pixel of each cut zone has identical class label.
2. according to the method described in claim 1, wherein step (2) carries out smoothly, being put down to the natural image after normalization
Sliding image It', carry out as follows：
(2a) sets glide filter window size to be 5 × 5, window iteration ends threshold epsilon1=0.001, space Gaussian kernel bandwidth σs
=1；
(2b) selection sliding window smoothing formula is as follows：
<mrow>
<msubsup>
<mi>y</mi>
<mi>i</mi>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msubsup>
<mo>=</mo>
<mfrac>
<mrow>
<munder>
<mo>&Sigma;</mo>
<mrow>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>&Element;</mo>
<msub>
<mi>N</mi>
<mi>i</mi>
</msub>
</mrow>
</munder>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>s</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<msub>
<mi>s</mi>
<mi>i</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<mrow>
<mn>2</mn>
<msubsup>
<mi>&sigma;</mi>
<mi>s</mi>
<mn>2</mn>
</msubsup>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<msubsup>
<mi>y</mi>
<mi>i</mi>
<mi>k</mi>
</msubsup>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<mrow>
<mn>2</mn>
<msubsup>
<mi>&sigma;</mi>
<mi>k</mi>
<mn>2</mn>
</msubsup>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
</mrow>
<mrow>
<munder>
<mo>&Sigma;</mo>
<mrow>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>&Element;</mo>
<msub>
<mi>N</mi>
<mi>i</mi>
</msub>
</mrow>
</munder>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>s</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<msub>
<mi>s</mi>
<mi>i</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<mrow>
<mn>2</mn>
<msubsup>
<mi>&sigma;</mi>
<mi>s</mi>
<mn>2</mn>
</msubsup>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<msubsup>
<mi>y</mi>
<mi>i</mi>
<mi>k</mi>
</msubsup>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<mrow>
<mn>2</mn>
<msubsup>
<mi>&sigma;</mi>
<mi>k</mi>
<mn>2</mn>
</msubsup>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mo><</mo>
<mn>5</mn>
<mo>></mo>
</mrow>
Wherein,Represent+1 iterative value of kth of the center pixel of i-th of sliding window, NiRepresent all pictures in sliding window
The set of element, siRepresent the space coordinate of sliding window central point pixel, sjRepresent that the neighborhood territory pixel space in sliding window is sat
Mark, σsRepresentation space constraint nucleus band is wide, σsIt is changeless during whole smothing filtering, σ is sets=1, σkRepresent to slide
Codomain nucleus band of the dynamic window in kth time iteration is wide, σkAs iterative process is what is constantly adjusted, its calculation formula is as follows：
<mrow>
<msub>
<mi>&sigma;</mi>
<mi>k</mi>
</msub>
<mo>=</mo>
<msqrt>
<mrow>
<mfrac>
<mn>1</mn>
<msub>
<mi>N</mi>
<mi>R</mi>
</msub>
</mfrac>
<munder>
<mo>&Sigma;</mo>
<mrow>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>&Element;</mo>
<msub>
<mi>N</mi>
<mi>i</mi>
</msub>
</mrow>
</munder>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<msubsup>
<mi>y</mi>
<mi>i</mi>
<mi>k</mi>
</msubsup>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
</msqrt>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mo><</mo>
<mn>6</mn>
<mo>></mo>
</mrow>
Wherein, NRThe quantity of pixel in sliding window is represented,Represent the kth time iteration of the center pixel of i-th of sliding window
Value, xjRepresent the pixel in sliding window, NiRepresent the set of all pixels point in sliding window；
(2c) sets first smooth pixel point subscript i=1；
(2d) sets window primary iteration number of times k=0,I=1,2 ..., N, N represent image slices vegetarian refreshments sum；
(2e) is according to formula<6>And formula<5>The corresponding self-adaptive kernel bandwidth σ of ith pixel is calculated respectivelykWith+1 iterative value of kth
(2f) judges that the front and rear iterative value twice of ith pixel point isWithIt is no to meetIf meeting,
I=i+1 is made, is performed (2g), otherwise, k=k+1 is made, returns (2e)；
(2g) judges whether i≤N sets up, if so, (2d) then is returned to, otherwise, calculating terminates.
3. according to the method described in claim 1, the use mean shift iterations operator wherein in step (4), initial with 64
Point is iterated search for starting point, carries out as follows：
(4a) sets average drifting codomain bandwidthMean shift iterations terminate threshold epsilon2=0.001；
(4b) mean shift iterations search calculation formula is as follows：
<mrow>
<msubsup>
<mi>y</mi>
<mi>m</mi>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msubsup>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<msub>
<mi>n</mi>
<mi>k</mi>
</msub>
</mfrac>
<munder>
<mo>&Sigma;</mo>
<mrow>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>&Element;</mo>
<msub>
<mi>S</mi>
<mi>h</mi>
</msub>
<mrow>
<mo>(</mo>
<msubsup>
<mi>y</mi>
<mi>m</mi>
<mi>k</mi>
</msubsup>
<mo>)</mo>
</mrow>
</mrow>
</munder>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mo><</mo>
<mn>1</mn>
<mo>></mo>
</mrow>
Wherein,Represent value of m-th of initial point after k mean shift iterations, m=1,2 ..., 64, Sh(yk) represent with
For the centre of sphere, the pixel point set surrounded using h as the ball of radius, nkRepresent the pixel point setThe number of middle pixel；
(4c) sets initial search point subscript m=1 of the average drifting in color space, makes iteration initial value
(4d) sets iterations k=0；
(4e) is according to formula<1>Calculate
(4f) judgesWhether set up：If so, then makeAnd m=m+1 is made, (4g) then is performed,
Otherwise k=k+1 is made, is returned (4e)；
(4g) judges whether m≤64 set up, if so, then make iteration initial valueReturn (4d), otherwise, terminate to calculate.
4. between according to the method described in claim 1, merging the convergence middle any two convergence points of point set V ' wherein in step (6)
Euclidean distance be less than threshold value h convergence point, obtain cluster centre set V ", carry out as follows：
(6a) sets left convergence point v 'pSubscript initial value p=1；
(6b) sets right convergence point v 'qSubscript initial value q=p+1；
(6c) judges whether q≤c sets up：If so, (6d) then is performed, otherwise, is performed (6f)；
(6d) calculates d=| | v 'p-v′q| |, judge whether d ＜ h set up：If so, then make v 'q=(v 'p+v′q)/2, c=c-1,
P=p+1, is returned (6b), otherwise, is performed (6e)；
(6e) makes q=q+1, judges whether q≤c sets up：If so, (6d) then is returned, otherwise, is performed (6f)；
(6f) makes p=p+1, judges whether p≤c-1 sets up：If so, (6b) then is returned, otherwise, terminates to calculate.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710434179.5A CN107301644B (en) | 2017-06-09 | 2017-06-09 | Natural image non-formaldehyde finishing method based on average drifting and fuzzy clustering |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710434179.5A CN107301644B (en) | 2017-06-09 | 2017-06-09 | Natural image non-formaldehyde finishing method based on average drifting and fuzzy clustering |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107301644A true CN107301644A (en) | 2017-10-27 |
| CN107301644B CN107301644B (en) | 2019-10-08 |
Family
ID=60134685
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710434179.5A Active CN107301644B (en) | 2017-06-09 | 2017-06-09 | Natural image non-formaldehyde finishing method based on average drifting and fuzzy clustering |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107301644B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108717069A (en) * | 2018-05-29 | 2018-10-30 | 电子科技大学 | A kind of high-pressure bottle thermal imaging imperfection detection method based on the segmentation of row variable step |
| CN109146894A (en) * | 2018-08-07 | 2019-01-04 | 庄朝尹 | A kind of model area dividing method of three-dimensional modeling |
| CN110717872A (en) * | 2019-10-08 | 2020-01-21 | 江西洪都航空工业集团有限责任公司 | Method and system for extracting characteristic points of V-shaped welding seam image under laser-assisted positioning |
| CN111931789A (en) * | 2020-07-28 | 2020-11-13 | 江苏大学 | Linear crop row extraction method suitable for different illumination, crop density and growth backgrounds |
| CN112308024A (en) * | 2020-11-23 | 2021-02-02 | 中国水利水电科学研究院 | Water body information extraction method |
| CN113409335A (en) * | 2021-06-22 | 2021-09-17 | 西安邮电大学 | Image segmentation method based on strong and weak joint semi-supervised intuitive fuzzy clustering |
| CN114332444A (en) * | 2021-12-27 | 2022-04-12 | 中国科学院光电技术研究所 | Complex starry sky background target identification method based on incremental drift clustering |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104299237A (en) * | 2014-10-20 | 2015-01-21 | 上海电机学院 | Image segmentation method converting unsupervised cluster into self-supervised classification |
| CN104751185A (en) * | 2015-04-08 | 2015-07-01 | 西安电子科技大学 | SAR image change detection method based on mean shift genetic clustering |
| CN106408580A (en) * | 2016-11-18 | 2017-02-15 | 南通大学 | Liver region extraction method based on fuzzy C mean and mean shift |
-
2017
- 2017-06-09 CN CN201710434179.5A patent/CN107301644B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104299237A (en) * | 2014-10-20 | 2015-01-21 | 上海电机学院 | Image segmentation method converting unsupervised cluster into self-supervised classification |
| CN104751185A (en) * | 2015-04-08 | 2015-07-01 | 西安电子科技大学 | SAR image change detection method based on mean shift genetic clustering |
| CN106408580A (en) * | 2016-11-18 | 2017-02-15 | 南通大学 | Liver region extraction method based on fuzzy C mean and mean shift |
Non-Patent Citations (1)
| Title |
|---|
| BO QU: "Research on image segmentation algorithm based on fuzzy clustering", 《FIFTH INTERNATIONAL CONFERENCE ON DIGITAL IMAGE PROCESSING(ICDIP 2013)》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108717069A (en) * | 2018-05-29 | 2018-10-30 | 电子科技大学 | A kind of high-pressure bottle thermal imaging imperfection detection method based on the segmentation of row variable step |
| CN108717069B (en) * | 2018-05-29 | 2020-08-11 | 电子科技大学 | High-pressure container thermal imaging defect detection method based on line variable step length segmentation |
| CN109146894A (en) * | 2018-08-07 | 2019-01-04 | 庄朝尹 | A kind of model area dividing method of three-dimensional modeling |
| CN110717872A (en) * | 2019-10-08 | 2020-01-21 | 江西洪都航空工业集团有限责任公司 | Method and system for extracting characteristic points of V-shaped welding seam image under laser-assisted positioning |
| CN111931789A (en) * | 2020-07-28 | 2020-11-13 | 江苏大学 | Linear crop row extraction method suitable for different illumination, crop density and growth backgrounds |
| CN112308024A (en) * | 2020-11-23 | 2021-02-02 | 中国水利水电科学研究院 | Water body information extraction method |
| CN113409335A (en) * | 2021-06-22 | 2021-09-17 | 西安邮电大学 | Image segmentation method based on strong and weak joint semi-supervised intuitive fuzzy clustering |
| CN113409335B (en) * | 2021-06-22 | 2023-04-07 | 西安邮电大学 | Image segmentation method based on strong and weak joint semi-supervised intuitive fuzzy clustering |
| CN114332444A (en) * | 2021-12-27 | 2022-04-12 | 中国科学院光电技术研究所 | Complex starry sky background target identification method based on incremental drift clustering |
| CN114332444B (en) * | 2021-12-27 | 2023-06-16 | 中国科学院光电技术研究所 | Complex star sky background target identification method based on incremental drift clustering |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107301644B (en) | 2019-10-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107301644B (en) | Natural image non-formaldehyde finishing method based on average drifting and fuzzy clustering | |
| Zhu et al. | A key volume mining deep framework for action recognition | |
| Ren et al. | Region-based saliency detection and its application in object recognition | |
| Xiao et al. | Multiple view semantic segmentation for street view images | |
| CN112101150B (en) | Multi-feature fusion pedestrian re-identification method based on orientation constraint | |
| CN112184752A (en) | Video target tracking method based on pyramid convolution | |
| CN108898145A (en) | A kind of image well-marked target detection method of combination deep learning | |
| CN105493078B (en) | Colored sketches picture search | |
| Rosenfeld et al. | Extracting foreground masks towards object recognition | |
| CN108389251A (en) | The full convolutional network threedimensional model dividing method of projection based on fusion various visual angles feature | |
| Sukanya et al. | A survey on object recognition methods | |
| CN108846404B (en) | Image significance detection method and device based on related constraint graph sorting | |
| Sheng et al. | Deep neural representation guided face sketch synthesis | |
| CN109101981B (en) | Loop detection method based on global image stripe code in streetscape scene | |
| Lu et al. | Localize me anywhere, anytime: a multi-task point-retrieval approach | |
| CN112712546A (en) | Target tracking method based on twin neural network | |
| Bindhu et al. | Hyperspectral image processing in internet of things model using clustering algorithm | |
| Zhang et al. | Fast moving pedestrian detection based on motion segmentation and new motion features | |
| CN108959379A (en) | A kind of image of clothing search method of view-based access control model marking area and cartographical sketching | |
| Andreetto et al. | Unsupervised learning of categorical segments in image collections | |
| Xue et al. | Real-world ISAR object recognition and relation discovery using deep relation graph learning | |
| Zhao et al. | Learning best views of 3D shapes from sketch contour | |
| Huang et al. | Graph cuts stereo matching based on patch-match and ground control points constraint | |
| Daryanto et al. | Survey: recent trends and techniques in image co-segmentation challenges, issues and its applications | |
| Wang et al. | Hypergraph based feature fusion for 3-D object retrieval |
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 |