CN106296654A - A kind of image superpixel dividing method keeping edge - Google Patents

A kind of image superpixel dividing method keeping edge Download PDF

Info

Publication number
CN106296654A
CN106296654A CN201610595744.1A CN201610595744A CN106296654A CN 106296654 A CN106296654 A CN 106296654A CN 201610595744 A CN201610595744 A CN 201610595744A CN 106296654 A CN106296654 A CN 106296654A
Authority
CN
China
Prior art keywords
node
super
pixel
image
seed
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
Application number
CN201610595744.1A
Other languages
Chinese (zh)
Inventor
陈雪锦
白宇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Science and Technology of China USTC
Original Assignee
University of Science and Technology of China USTC
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by University of Science and Technology of China USTC filed Critical University of Science and Technology of China USTC
Priority to CN201610595744.1A priority Critical patent/CN106296654A/en
Publication of CN106296654A publication Critical patent/CN106296654A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of image superpixel dividing method keeping edge, including: the image of input is converted to non-directed graph structure, the most therefrom extract the minimum spanning tree MST of image, image is carried out monolayer super-pixel segmentation: the basic tree distance for measurement image space information is uniformity criterion, MST uniformly takes seed points, as class central point；For weighing the weight tree distance of image color information as diversity criterion, with the strategy of arest neighbors, the remaining node removed in MST outside class central point is clustered；Image is carried out high-rise super-pixel segmentation: from the result of monolayer super-pixel segmentation, extract a new MST, as the input of next layer of super-pixel segmentation, and repeat monolayer super-pixel segmentation step；Meanwhile, along with the increase of the number of plies gradually reduces dividing number, thus desired super-pixel dividing number is generated after last layer of super-pixel segmentation.The program can protect image border, meanwhile, also improves treatment effeciency.

Description

A kind of image superpixel dividing method keeping edge
Technical field
The present invention relates to image analysis technology field, particularly relate to a kind of image superpixel dividing method keeping edge.
Background technology
The data expansion that the rise of computer and the Internet brings is inevitable, and the update of the equipment such as photographing unit makes The size of picture is increasing, deals with the most more and more difficult.If going of putting pixel-by-pixel is analyzed, the information that can cause superfluous Waste that is remaining and that calculate.If there is a kind of method, in advance pixel classifications close to color similarity and position to together, as One entirety, as successive image analyze in an elementary cell, this will be significantly reduced graphical analysis amount of calculation and The operation time.
Super-pixel is split, it is simply that such a preprocess method, if its objective is to divide the image in advance color similarity Dry region, reduces the redundancy between pixel.After image does super-pixel segmentation, we can be each piece after segmentation Cut zone processes as a whole removing, replaces analysis the most pixel-by-pixel with the analysis of block-by-block, thus greatly reduces The quantity of feature, has saved the time calculated；As it is shown in figure 1, left side is input picture, centre is traditional super-pixel segmentation side Input picture is split by formula, and right side is segmentation result.But, traditional super-pixel partitioning scheme can not keep figure completely The information of picture, when super-pixel dividing number reduces, image information meeting more or less is lost.So, this research purpose exists In by the least dividing number, the detailed information of holding artwork as well as possible.
Super-pixel method is largely divided into two kinds: algorithm based on figure segmentation and algorithm based on gradient.For based on figure point The algorithm cut, they often first obtain a non-directed graph structure from image, then split energy equation by optimizing parent map, no Disconnected splits non-directed graph, until generating last super-pixel segmentation result.Different algorithms has different figures to divide Cut energy equation definition, and different optimization methods.But, even simplest energy equation, in solution procedure also It is the problem of a NP-hard, and goes to solve at whole global image, need to take the highest time complexity.Right In method based on gradient, the most first choose a number of pixel as seed points, then utilize other pixel to seed Other pixel is clustered by the distance (diversity) of point.But, be no matter use based on color and coordinate European away from From, it being also based on the geodesic distance in path, they are unsatisfactory to the protection of marginal information in image.And, generally require many Secondary identical iterative computation, the effect that super-pixel segmentation result could be converged on protection image border, which results in Algorithm operational efficiency based on gradient is the lowest.
Summary of the invention
It is an object of the invention to provide a kind of image superpixel dividing method keeping edge, it is possible to protection image border, Meanwhile, treatment effeciency is also improved.
It is an object of the invention to be achieved through the following technical solutions:
A kind of image superpixel dividing method keeping edge, including:
The image of input is converted to non-directed graph structure, and the method for recycling linear complexity is extracted from non-directed graph structure Publish picture the minimum spanning tree MST of picture, and defines in MST for weighing the basic tree distance of image space information and for weighing figure Weight tree distance as colouring information；
Image carries out monolayer super-pixel segmentation, and its step includes: with basic tree distance for uniformity criterion, Seed points is uniformly taked, as class central point on MST；Using weight tree distance as diversity criterion, by the plan of arest neighbors Slightly the remaining node outside removing class central point in MST is clustered；
Image is carried out high-rise super-pixel segmentation: from the result of monolayer super-pixel segmentation, extract a new MST, make The input split for next layer of super-pixel, and repeat the step of monolayer super-pixel segmentation；Meanwhile, along with the number of plies increase gradually Reduce dividing number, thus generate desired super-pixel dividing number after last layer of super-pixel segmentation.
Described the image of input be converted to non-directed graph structure include:
The node being non-directed graph structure with the pixel of input picture, with four neighborhood relationships between pixel for non-directed graph structure Limit, then non-directed graph representation is G=(V, E)；V therein is node set, and E is limit set；
The weight on each limit is that the RGB channel maximum color of adjacent two nodes of four neighborhoods is poor, is expressed as: e (pi,pi+1) =maxe∈{R,G,B}||Ic(pi)-Ic(pi+1)||；Wherein, e (pi,pi+1) it is adjacent node piWith pi+1Colour-difference, IcFor color Value.
Described basic tree distance is designated as DS(p q), represents the node p and the node q length in MST upper pathway；If node p It is adjacent node then D with node qS(p, q)=1；
Described weight tree distance is designated as DC(p q), represents node p and node q the asking of all limits weight on the path of MST With；If p and q is adjacent node, DS(p, q)=e (p, q)；Wherein, (p is q) node p and the colour-difference of node q to e.
Described with basic tree distance for uniformity criterion, the step uniformly taking seed points on MST is:
Initialize: preset sampling site distance DSeed, create two empty queue LSeedAnd LCandidateStore the kind adopted respectively Son point and undetermined alternative seed points, and the root node of MST is put into LCandidateIn；
Determine and adopt seed points and newly-increased alternative seed points: from LCandidateOne node p of middle taking-up, if LSeedNon-NULL, Then check whether node p meets and LSeedIn already present seed points distance less than setting value LSeedCondition；If it is full Foot, then insert L using node p as seed pointsSeedIn queue, and the depth-first traversal of the degree of depth that fixes with node p for starting point, New alternative seed points is looked in MST；
During depth-first traversal, basic tree distance D of any traversing nodes q to start node pS(p, q) be with The increase of the degree of depth and progressively accumulation calculating；Once meet DS(p, q) ＞ DSeed, then node q is put into alternative seed points team Row LCandidateIn, and stop at the depth-first traversal of node q place branch；By carrying out depth-first traversal with upper type Afterwards, all alternative seed points queue L is all put into node p for the alternative seed points that starting point findsCandidateIn；
Again from LCandidateTake out an alternative seed points repeat above-mentioned determine adopt seed points and newly-increased alternative seed points Step, until LCandidateFor empty set.
Described using weight tree distance as diversity criterion, with the strategy of arest neighbors to MST removes class central point Outside remaining node do cluster and include:
Giving a label value for each class central point, label value is equal to class central point at seed points queue LSeedIn sequence Column number value；
Set clustering distance DCluster(p, q), this clustering distance DCluster(p, q) utilize a pair be bordered by a little to weight tree Distance calculates:
DCluster(p, q)=Σ max (DC(pa, pb),1)；
Wherein, pa, pbBe connect node p and q path on be bordered by for any pair a little right, DC(pa,pb) for being bordered by a little to pa With pbWeight tree distance；
With clustering distance DCluster(p q) is the nearest neighbouring rule of retrochromism criterion between pixel；For MST In each node p being not yet classified, be classified into from its nearest class central node, i.e. the apoplexy due to endogenous wind nearest from node p The label value of heart point is given to node p:
L ( p ) = L ( arg min q ∈ L Seed D Cluster ( p , q ) ) ;
Wherein, LSeedFor setting value.
Described from the result of monolayer super-pixel segmentation, extract a new MST, as the segmentation of next layer of super-pixel Input includes:
Assume that the width of image I and high W and H, image I of being respectively are divided into K0H region of=W, each cut section The pixel quantity N that territory s comprises0(p)=1；
Then after the segmentation of r-1 layer super-pixel, image is divided into Kr-1Individual region, the pixel that each cut zone s comprises Quantity is Nr-1(s)；
Splitting for r layer super-pixel, in the result split r-1 layer super-pixel, each piece of cut zone s is considered as one Individual node, builds limit between adjacent region and connects each node, thus construct a new non-directed graph structure, limit Weight is that the average color of adjacent two cut zone is poor:
e r ( s , s ′ ) = | | I ‾ r - 1 ( s ) - I ‾ r - 1 ( s ′ ) | |
Wherein,It is respectively the average color of region s Yu s' in r-1 layer super-pixel segmentation result；
The input that the MST made new advances is split as r layer super-pixel is extracted again from new non-directed graph structure.
The described increase along with the number of plies gradually reduces dividing number, thus generates the phase after last layer of super-pixel segmentation The super-pixel dividing number hoped includes:
The quantity of super-pixel segmentation is along with sampling site distance DSeedIncrease and monotone decreasing；To any one layer of r, define DSeed =kdr-1, wherein, d controls DSeedGrowth rate, i.e. super-pixel segmentation minimizing speed；K controls DSeedInitial size, i.e. Dividing number after splitting for the first time from artwork；
Assume that desired super-pixel dividing number isWhen the number of plies increases, DSeedValue increase, then there is a certain layer l MakeAnd
As seen from the above technical solution provided by the invention, the super-pixel segmentation protecting limit based on minimum spanning tree is calculated Method first determines monolayer dividing method fast and effectively, then improves final super-pixel segmentation effect by multi-layer segmentation.
Accompanying drawing explanation
In order to be illustrated more clearly that the technical scheme of the embodiment of the present invention, required use in embodiment being described below Accompanying drawing be briefly described, it should be apparent that, below describe in accompanying drawing be only some embodiments of the present invention, for this From the point of view of the those of ordinary skill in field, on the premise of not paying creative work, it is also possible to obtain other according to these accompanying drawings Accompanying drawing.
The traditional images super-pixel segmentation schematic diagram that Fig. 1 provides for background of invention；
The flow chart of a kind of image superpixel dividing method keeping edge that Fig. 2 provides for the embodiment of the present invention.
Detailed description of the invention
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete Ground describes, it is clear that described embodiment is only a part of embodiment of the present invention rather than whole embodiments.Based on this Inventive embodiment, the every other enforcement that those of ordinary skill in the art are obtained under not making creative work premise Example, broadly falls into protection scope of the present invention.
The embodiment of the present invention provides a kind of image superpixel dividing method keeping edge, as in figure 2 it is shown, it mainly wraps Include:
Step 21, the image of input being converted to non-directed graph structure, the method for recycling linear complexity is tied from non-directed graph Structure extracts the minimum spanning tree MST of image, and defines in MST for weighing basic tree distance and the use of image space information In the weight tree distance weighing image color information.
In the embodiment of the present invention, image being considered as a non-directed graph structure, node is made up of pixel, while between by pixel Four neighborhood relationships compositions, the weight on limit is the colour-difference of adjacent pixels.In the middle of non-directed graph, a MST can be extracted (minimum spanning tree, minimum spanning tree), because the limit (often in marginal portion) of big weight is during contributing Can be removed, so minimum spanning tree structure can highlight the marginal information in image.Tree distance, be defined in MST away from From, the path being connected between two nodes on MST determine.Owing to path is to be connected by the limit that a series of weights are minimum Connect, it is possible to avoid the big limit of weights (the most removed limit) so that the tree distance being defined on MST can preferably be retouched Stating edge two and survey the diversity between (zones of different) pixel, protecting limit property image smoothing, stereoscopic vision coupling field obtains It is widely applied.
Step 22, image carrying out monolayer super-pixel segmentation, its step includes: weigh mark with basic tree distance for uniformity Standard, uniformly takes seed points, as class central point on MST；Using weight tree distance as diversity criterion, use arest neighbors Strategy cluster MST removes the remaining node outside class central point.
In the embodiment of the present invention, the segmentation of monolayer super-pixel is broadly divided into sampling site and two steps of cluster, in sampling site step, Use basic tree distance as criterion, by a kind of method being similar to Poisson sampling site, MST uniformly take seed points, As class central point.
In sorting procedure, use arest neighbors strategy based on weight tree distance, according to the class central point selected to surplus in MST Remaining tree node clusters.In cluster process, when calculating the attached node of each class central node, it is all the searching of local, So the time complexity of monolayer segmentation and image size are linear.
Step 23, image is carried out the segmentation of high-rise super-pixel: extract from the result of monolayer super-pixel segmentation one new MST, as the input of next layer of super-pixel segmentation, and repeat the step of monolayer super-pixel segmentation；Meanwhile, along with the increasing of the number of plies Add and gradually reduce dividing number, thus generate desired super-pixel dividing number after last layer of super-pixel segmentation.
Super-pixel owing to being generated by monolayer segmentation is split, it is easy to affected by image texture.Because sampling site away from From excessive so that the meeting of weight tree distance is cumulative the greatlyyest, the colour-difference on the most real border.So, high level is proposed The method of segmentation, changes a step and obtains the strategy of given super-pixel dividing number, but minimizing super-pixel of going successively is split Quantity, and each piece of region of gradually smooth segmentation during dividing number reduces, arrive the mesh of suppression texture noise 's.In splitting at each layer, use the method similar with monolayer segmentation, the graph structure of this layer is split.And current layer Each piece of region of segmentation result, all using a figure node as next layer of graph structure.Due to super-pixel segmentation quantity with The number of plies increase and and index decreased, so, multi-layer segmentation algorithm mainly concentrate front which floor, therefore, time complexity is also Linear with the size of image.
The multilamellar super-pixel partitioning algorithm that the embodiment of the present invention proposes, is to count on the CPU of 4.0HZ in Intel's I7 dominant frequency Calculating, the operation time is about 25FPS.Being significantly faster than other existing popular super-pixel partitioning algorithm will be fast, and also Have the effect preferably protecting image border.
In order to make it easy to understand, elaborate below for each step above-mentioned.
1, extract MST (minimum spanning tree), and define two kinds of tree distances.
In the embodiment of the present invention, the image of input is converted to non-directed graph structure；Wherein, with the pixel of input picture as nothing To the node of graph structure, the limit being non-directed graph structure with four neighborhood relationships between pixel, then non-directed graph representation is G= (V,E)；V therein is node set, and E is limit set；
The weight on each limit is the colour-difference of adjacent node, can be defined as the RGB channel of adjacent two pixels of four neighborhoods Maximum color is poor, is expressed as: e (pi,pi+1)=maxe∈{R,G,B}||Ic(pi)-Ic(pi+1)||；Wherein, e (pi,pi+1) it is adjacent Node piWith pi+1Colour-difference, IcFor color value.
Afterwards, we can extract the MST of image from non-directed graph structure by the method for linear complexity.
In the embodiment of the present invention, in MST, the present invention implements based on two kinds of tree distances: substantially sets distance and (is used for weighing figure Image space information) and weight tree distance (being used for weighing image color information).
Described basic tree distance is designated as DS(p q), represents the node p and the node q length in MST upper pathway；If node p It is adjacent node then D with node qS(p, q)=1；Basic tree distance reflects two nodes on image to a certain extent The alternate position spike opposite sex.
Described weight tree distance is designated as DC(p q), represents node p and node q the asking of all limits weight on the path of MST With；If p and q is adjacent node, DS(p, q)=e (p, q)；Wherein, (p is q) node p and the colour-difference of node q to e.Add Quan Shu distance reflects two nodes retrochromism on image.
2, image is carried out monolayer super-pixel segmentation.
Image is carried out monolayer super-pixel segmentation, is divided into sampling site and cluster.In sampling site, with basic tree distance as uniformity Criterion, uniform sampling site on MST；In cluster, using weight tree distance as diversity criterion, by the plan of nearest neck Slightly other node on MST is clustered.Detailed process is as follows:
1) sampling site: in monolayer super-pixel partitioning algorithm, in order to allow seed points (i.e. class central point) be evenly distributed in In little spanning tree, the embodiment of the present invention uses a kind of method being similar to Poisson sampling site to take seed in tree construction uniformly Point.Owing to basic tree distance the most more can embody the spatial information of image, so, we are with basic tree distance for weighing all The standard of even property, takes seed points in tree construction uniformly.Assume DSeedFor sampling site distance (during being similar to Poisson sampling site The definition of sampling site radius r and effect), by a series of based on specific sampling site distance DSeedIterative computation, can by with Lower method takes seed points from tree construction uniformly:
Initialize: preset sampling site distance DSeed, create two empty queue LSeedAnd LCandidateStore the kind adopted respectively Son point and undetermined alternative seed points, and the root node of MST is put into LCandidateIn；
Determine and adopt seed points and newly-increased alternative seed points: from LCandidateOne node p of middle taking-up, if LSeedNon-NULL, Then check whether node p meets and LSeedIn already present seed points distance less than setting value LSeedCondition；If it is full Foot, then insert L using node p as seed pointsSeedIn queue, and the depth-first traversal of the degree of depth that fixes with node p for starting point, New alternative seed points is looked in MST；
During depth-first traversal, basic tree distance D of any traversing nodes q to start node pS(p, q) be with The increase of the degree of depth and progressively accumulation calculating；Once meet DS(p, q) ＞ DSeed, then node q is put into alternative seed points team Row LCandidateIn, and stop at the depth-first traversal of node q place branch；By carrying out depth-first traversal with upper type Afterwards, all alternative seed points queue L is all put into node p for the alternative seed points that starting point findsCandidateIn；
Again from LCandidateTake out an alternative seed points repeat above-mentioned determine adopt seed points and newly-increased alternative seed points Step, until LCandidateFor empty set.
2) cluster: give a label value for each class central point, label value is equal to class central point in seed points queue LSeedIn sequence number value；
Set clustering distance DCluster(p, q), this clustering distance DCluster(p, q) utilize a pair be bordered by a little to weight tree Distance calculates:
DCluster(p, q)=∑ max (DC(pa, pb),1)；
Wherein, pa, pbBe connect node p and q path on be bordered by for any pair a little right, DC(pa,pb) for being bordered by a little to pa With pbWeight tree distance；
Above-mentioned clustering distance DCluster(p, q) can overcome weight tree distance is zero to lead at image flat site accumulated value Cause cannot be by the shortcoming of arest neighbors classification.
With clustering distance DCluster(p q) is the nearest neighbouring rule of retrochromism criterion between pixel；For MST In each node p being not yet classified, be classified into from its nearest class central node, i.e. the apoplexy due to endogenous wind nearest from node p The label value of heart point is given to node p:
L ( p ) = L ( arg min q ∈ L Seed D Cluster ( p , q ) ) ;
Wherein, LSeedFor setting value, L is corresponding label value.
Additionally, due to the basic tree distance of two nodes and weight tree distance are accumulation calculating and based on phase in MST Same path, it is possible to the depth-first traversal in same constant depth calculates D simultaneouslyS(p, q) (for sampling site) with DC (p q) (is used for clustering).
3, image is carried out multilamellar super-pixel segmentation.
If apart from each other between sampled point, weight tree distance may add up excessive in image texture region, more than true Positive border color is poor and causes the arest neighbors classification of mistake in monolayer segmentation.So the algorithm extension that monolayer is split by the present invention In multi-layer segmentation, each layer of segmentation use the algorithm of our monolayer, and utilize the segmentation result smoothed image of current layer, Input as next layer.In multilamellar super-pixel segmenting structure, invention replaces a direct step and calculate desired segmentation Quantity, and the increase being as the number of plies gradually reduces dividing number, and generate desired super-pixel segmentation number at last layer Amount.
Because the method that the segmentation of each layer is all based on monolayer segmentation, sampling site and cluster is i.e. utilized tree construction to be carried out point Cut.So expand to multi-layer segmentation focuses on how connecting relation between layers, the most how from the segmentation of current layer As a result, a new non-directed graph structure is extracted, as the input of next layer；Detailed process is as follows:
A, the width assuming image I and height are respectively W and H, image I and are divided into K0H region of=W, each segmentation The pixel quantity N that region s comprises0(p)=1；
After b, then r-1 layer super-pixel segmentation, image is divided into Kr-1Individual region, any one this layer of cut zone s bag The pixel quantity contained is Nr-1(s)；
C, for r layer super-pixel split, by r-1 layer super-pixel split result in each piece of cut zone s be considered as One node, builds limit between adjacent region and connects each node, thus construct a new non-directed graph structure, limit The average color that weight is adjacent two cut zone poor:
e r ( s , s ′ ) = | | I ‾ r - 1 ( s ) - I ‾ r - 1 ( s ′ ) | |
Wherein,It is respectively the average color of region s Yu s' in r-1 layer super-pixel segmentation result.
D, from new non-directed graph structure, extract the input that the MST made new advances is split as r layer super-pixel again.
Above-mentioned a～d process is it is to be understood that according to the segmentation result of r-1 layer, take single face to each piece of cut zone Color, color value is the average color in this region.Then, with each piece of region as node, between adjacent region, build limit come Connect each node, obtain new non-directed graph, then from new non-directed graph, extract the MST made new advances, as the input of r layer.
In the multi-layer segmentation structure of the embodiment of the present invention, the quantity of super-pixel segmentation is along with sampling site distance DSeedIncrease And monotone decreasing；To any one layer of r, define DSeed=kdr-1, wherein, d controls DSeedGrowth rate, i.e. super-pixel segmentation Reduce speed；K controls DSeedInitial size, i.e. dividing number after for the first time segmentation from artwork；Assume desired super picture Element dividing number isWhen the number of plies increases, DSeedValue increase, then there is a certain layer l and makeAnd
Additionally, due in each layer of non-directed graph in multi-layer segmentation of the present invention the quantity of node be as the number of plies increase and Index decreased.So, the main amount of calculation of multi-layer segmentation concentrates several leading layer in the frame, in complexity and image pixel Quantity is linear.
The framework of the multi-layer segmentation proposed in the embodiment of the present invention, gradually decreases the number of super-pixel segmentation with number of plies increase Amount, and image is done smooth in every layer of segmentation, reduce the impact of texture cumulative errors.The complexity of every layer of segmentation increases with the number of plies Add index decreased, therefore Riming time of algorithm is concentrated mainly on the front in which floor of multi-layer segmentation, and image pixel is linear.? Experiment on Berkeley image partition data storehouse may certify that, compares existing epidemic algorithms, and it is super that the embodiment of the present invention provides Pixel segmentation has preferably protects limit property, and time efficiency is higher.
Through the above description of the embodiments, those skilled in the art it can be understood that to above-described embodiment can To be realized by software, it is also possible to the mode adding necessary general hardware platform by software realizes.Based on such understanding, The technical scheme of above-described embodiment can embody with the form of software product, this software product can be stored in one non-easily The property lost storage medium (can be CD-ROM, USB flash disk, portable hard drive etc.) in, including some instructions with so that a computer sets Standby (can be personal computer, server, or the network equipment etc.) performs the method described in each embodiment of the present invention.
The above, the only present invention preferably detailed description of the invention, but protection scope of the present invention is not limited thereto, Any those familiar with the art in the technical scope of present disclosure, the change that can readily occur in or replacement, All should contain within protection scope of the present invention.Therefore, protection scope of the present invention should be with the protection model of claims Enclose and be as the criterion.

Claims (7)

1. the image superpixel dividing method keeping edge, it is characterised in that including:
The image of input is converted to non-directed graph structure, and the method for recycling linear complexity is extracted from non-directed graph structure and is published picture The minimum spanning tree MST of picture, and define in MST for weighing the basic tree distance of image space information and for weighing image face The weight tree distance of color information；
Image carries out monolayer super-pixel segmentation, and its step includes: with basic tree distance for uniformity criterion, on MST Uniformly take seed points, as class central point；Using weight tree distance as diversity criterion, right with the strategy of arest neighbors MST removes the remaining node outside class central point cluster；
Image is carried out the segmentation of high-rise super-pixel: from the result of monolayer super-pixel segmentation, extract a new MST, as under The input of one layer of super-pixel segmentation, and repeat the step of monolayer super-pixel segmentation；Meanwhile, gradually reduce along with the increase of the number of plies Dividing number, thus generate desired super-pixel dividing number after last layer of super-pixel segmentation.
A kind of image superpixel dividing method keeping edge the most according to claim 1, it is characterised in that described by defeated The image entered is converted to non-directed graph structure and includes:
The node being non-directed graph structure with the pixel of input picture, with four neighborhood relationships between pixel for non-directed graph structure Limit, then non-directed graph representation is G=(V, E)；V therein is node set, and E is limit set；
The weight on each limit is that the RGB channel maximum color of adjacent two nodes of four neighborhoods is poor, is expressed as: e (pi,pi+1)= maxe∈{R,G,B}||Ic(pi)-Ic(pi+1)||；Wherein, e (pi,pi+1) it is adjacent node piWith pi+1Colour-difference, IcFor color value.
A kind of image superpixel dividing method keeping edge the most according to claim 1, it is characterised in that
Described basic tree distance is designated as DS(p q), represents the node p and the node q length in MST upper pathway；If node p and knot Point q is adjacent node then DS(p, q)=1；
Described weight tree distance is designated as DC(p q), represents node p and node q summation of all limits weight on the path of MST；As Really p and q is adjacent node then DS(p, q)=e (p, q)；Wherein, (p is q) node p and the colour-difference of node q to e.
A kind of image superpixel dividing method keeping edge the most according to claim 1, it is characterised in that described with base This tree distance is uniformity criterion, and the step uniformly taking seed points on MST is:
Initialize: preset sampling site distance DSeed, create two empty queue LSeedAnd LCandidateStore the seed points adopted respectively With undetermined alternative seed points, and the root node of MST is put into LCandidateIn；
Determine and adopt seed points and newly-increased alternative seed points: from LCandidateOne node p of middle taking-up, if LSeedNon-NULL, then examine Whether a p that comes to an end meets and LSeedIn already present seed points distance less than setting value LSeedCondition；If it is satisfied, then Node p is inserted L as seed pointsSeedIn queue, and the depth-first traversal of the degree of depth that fixes with node p for starting point, at MST In look for new alternative seed points；
During depth-first traversal, basic tree distance D of any traversing nodes q to start node pS(p q) is as the degree of depth Increase and progressively accumulation calculating；Once meet DS(p, q) ＞ DSeed, then node q is put into alternative seed points queue LCandidateIn, and stop at the depth-first traversal of node q place branch；By with upper type carry out depth-first traversal it After, all put into alternative seed points queue L with node p for the alternative seed points that starting point findsCandidateIn；
Again from LCandidateTake out an alternative seed points and repeat the above-mentioned step determining and adopting seed points with newly-increased alternative seed points Suddenly, until LCandidateFor empty set.
A kind of image superpixel dividing method keeping edge the most according to claim 1, it is characterised in that described to add Quan Shu distance is as diversity criterion, with the strategy of arest neighbors to the remaining node outside removing class central point in MST Do cluster to include:
Giving a label value for each class central point, label value is equal to class central point at seed points queue LSeedIn serial number Value；
Set clustering distance DCluster(p, q), this clustering distance DCluster(p, q) utilize a pair be bordered by a little to weight tree distance Calculate:
DCluster(p, q)=∑ max (DC(pa,pb),1)；
Wherein, pa, pbBe connect node p and q path on be bordered by for any pair a little right, DC(pa,pb) for being bordered by a little to paWith pb's Weight tree distance；
With clustering distance DCluster(p q) is the nearest neighbouring rule of retrochromism criterion between pixel；For every in MST The individual node p being not yet classified, is classified into from its nearest class central node, i.e. the class central point nearest from node p Label value be given to node p:
L ( p ) = L ( argmin q ∈ L S e e d D C l u s t e r ( p , q ) ) ;
Wherein, LSeedFor setting value.
A kind of image superpixel dividing method keeping edge the most according to claim 1, it is characterised in that described from list Extracting a new MST in the result of layer super-pixel segmentation, the input as next layer of super-pixel segmentation includes:
Assume that the width of image I and high W and H, image I of being respectively are divided into K0H region of=W, each cut zone s bag The pixel quantity N contained0(p)=1；
Then after the segmentation of r-1 layer super-pixel, image is divided into Kr-1Individual region, the pixel quantity that each cut zone s comprises For Nr-1(s)；
Splitting for r layer super-pixel, in the result split r-1 layer super-pixel, each piece of cut zone s is considered as a knot Point, builds limit between adjacent region and connects each node, thus construct a new non-directed graph structure, the weight on limit Average color for adjacent two cut zone is poor:
e r ( s , s ′ ) = | | I ‾ r - 1 ( s ) - I ‾ r - 1 ( s ′ ) | |
Wherein,It is respectively the average color of region s Yu s' in r-1 layer super-pixel segmentation result；
The input that the MST made new advances is split as r layer super-pixel is extracted again from new non-directed graph structure.
A kind of image superpixel dividing method keeping edge the most according to claim 1, it is characterised in that described along with The increase of the number of plies gradually reduces dividing number, thus generates desired super-pixel segmentation number after last layer of super-pixel segmentation Amount includes:
The quantity of super-pixel segmentation is along with sampling site distance DSeedIncrease and monotone decreasing；To any one layer of r, define DSeed=kdr -1, wherein, d controls DSeedGrowth rate, i.e. super-pixel segmentation minimizing speed；K controls DSeedInitial size, i.e. from former Dividing number after splitting for the first time in figure；
Assume that desired super-pixel dividing number isWhen the number of plies increases, DSeedValue increase, then there is a certain layer l and makeAnd
CN201610595744.1A 2016-07-26 2016-07-26 A kind of image superpixel dividing method keeping edge Pending CN106296654A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610595744.1A CN106296654A (en) 2016-07-26 2016-07-26 A kind of image superpixel dividing method keeping edge

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610595744.1A CN106296654A (en) 2016-07-26 2016-07-26 A kind of image superpixel dividing method keeping edge

Publications (1)

Publication Number Publication Date
CN106296654A true CN106296654A (en) 2017-01-04

Family

ID=57652340

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610595744.1A Pending CN106296654A (en) 2016-07-26 2016-07-26 A kind of image superpixel dividing method keeping edge

Country Status (1)

Country Link
CN (1) CN106296654A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106919950A (en) * 2017-01-22 2017-07-04 山东大学 Probability density weights the brain MR image segmentation of geodesic distance
CN107492101A (en) * 2017-09-07 2017-12-19 四川大学 Multi-modal rhinopharyngeal neoplasm partitioning algorithm based on the optimal figure of self-adaptive construction
CN108876789A (en) * 2018-06-15 2018-11-23 南方医科大学 A kind of sequential chart segmentation method combined based on super-pixel and neighborhood block feature
CN113663331A (en) * 2021-08-03 2021-11-19 网易（杭州）网络有限公司 Identification generation method and device, nonvolatile storage medium and electronic device
CN114677384A (en) * 2022-03-13 2022-06-28 江苏神州新能源电力有限公司 Solar cell coating defect detection method

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102800093A (en) * 2012-07-12 2012-11-28 西安电子科技大学 Multi-target remote sensing image segmentation method based on decomposition
CN105469369A (en) * 2015-11-27 2016-04-06 中国科学院自动化研究所 Digital image filtering method and system based on segmented image

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102800093A (en) * 2012-07-12 2012-11-28 西安电子科技大学 Multi-target remote sensing image segmentation method based on decomposition
CN105469369A (en) * 2015-11-27 2016-04-06 中国科学院自动化研究所 Digital image filtering method and system based on segmented image

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
Y.BAI,X.J.CHEN: ""Efficient structure-preserving Superpixel Segmentation Based on Minimum Spanning Tree"", 《2016 IEEE INTERNATIONAL CONFERENCE ON MULTIMEDIA AND EXPO (ICME)》 *

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106919950A (en) * 2017-01-22 2017-07-04 山东大学 Probability density weights the brain MR image segmentation of geodesic distance
CN106919950B (en) * 2017-01-22 2019-10-25 山东大学 The brain MR image segmentation of probability density weighting geodesic distance
CN107492101A (en) * 2017-09-07 2017-12-19 四川大学 Multi-modal rhinopharyngeal neoplasm partitioning algorithm based on the optimal figure of self-adaptive construction
CN108876789A (en) * 2018-06-15 2018-11-23 南方医科大学 A kind of sequential chart segmentation method combined based on super-pixel and neighborhood block feature
CN108876789B (en) * 2018-06-15 2022-03-25 南方医科大学 Continuous graph cutting method based on combination of super-pixel and neighborhood block features
CN113663331A (en) * 2021-08-03 2021-11-19 网易（杭州）网络有限公司 Identification generation method and device, nonvolatile storage medium and electronic device
CN114677384A (en) * 2022-03-13 2022-06-28 江苏神州新能源电力有限公司 Solar cell coating defect detection method
CN114677384B (en) * 2022-03-13 2022-12-02 江苏神州新能源电力有限公司 Solar cell coating defect detection method

Similar Documents

Publication Publication Date Title
CN106296654A (en) A kind of image superpixel dividing method keeping edge
Veksler Stereo correspondence by dynamic programming on a tree
US8428363B2 (en) Method for segmenting images using superpixels and entropy rate clustering
Couprie et al. Power watersheds: A new image segmentation framework extending graph cuts, random walker and optimal spanning forest
US8913830B2 (en) Multilevel image segmentation
US8411952B2 (en) Method for segmenting an image using constrained graph partitioning of watershed adjacency graphs
WO2018227800A1 (en) Neural network training method and device
US8712154B2 (en) Image processing system and method
Miranda et al. Oriented image foresting transform segmentation by seed competition
US7978913B2 (en) Image segmentation based on processing of a graph
CN109035194B (en) Blood vessel extraction method and device
CN110032682A (en) A kind of information recommendation list generation method, device and equipment
CN110069989A (en) Face image processing process and device, computer readable storage medium
CN109948575A (en) Eyeball dividing method in ultrasound image
Buyssens et al. Eikonal-based vertices growing and iterative seeding for efficient graph-based segmentation
CN110084181A (en) A kind of remote sensing images Ship Target Detection method based on sparse MobileNetV2 network
Leitner et al. A computational study of exact approaches for the bi-objective prize-collecting steiner tree problem
CN105809696A (en) Automatic positioning and segmentation method for adrenal gland in sequence CT image
CN107403465B (en) Urban scene segmentation plane reconstruction method based on structure prior and deep learning
KR102046240B1 (en) Apparatus and method for lesion screening
CN112446951B (en) Three-dimensional reconstruction method, three-dimensional reconstruction device, electronic equipment and computer storage medium
Geetha et al. An improved method for segmentation of point cloud using minimum spanning tree
Siva et al. Grid seams: A fast superpixel algorithm for real-time applications
Shibata et al. Image inpainting based on probabilistic structure estimation
CN116797830A (en) Image risk classification method and device based on YOLOv7

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20170104

RJ01 Rejection of invention patent application after publication