CN106340020A - Super-pixel segmentation method and device - Google Patents

Abstract

The invention provides a super-pixel segmentation method and device. The method comprises the steps that N pixel points use multidimensional space values, and parameters for controlling clustering tightness are preset; Euclidean distance between every two pixel points in N pixel points is calculated; the initial center points of K clusters are selected among N pixel points through the preset minimum Euclidean distance and the center point probability value; a circular ring inequation is used; the nearest center point of other pixel points is determined among K initial center points; and the pixel points which belong to the same center point are identified as the same super-pixel to realize K super-pixel segmentation. According to the embodiment provided by the invention, compactness parameters are used; the super-pixel compactness can be directly controlled; the center points of K super-pixels are determined through the circular ring inequation; the number of times of inner loop of a K-means clustering method is greatly reduced; the super-pixel segmentation speed is accelerated; and the problems of slow segmentation speed and poor compactness in the prior art are solved.

Description

Superpixel segmentation method and device

Technical field

The present embodiments relate to image processing techniquess, more particularly, to a kind of superpixel segmentation method and device.

Background technology

The sensuously consistent pixel clusters in ability domain representation digital picture of super-pixel (superpixel), by Multiple pixel compositions, typically appear as identical or approximate color and brightness.Super-pixel utilizes pixel and picture Between element, the similarity degree of feature is grouped to pixel, thus obtaining the redundancy of image, reduces The complexity of successive image process task.

Super-pixel segmentation technology can be used for the application such as target recognition, image segmentation, three-dimensional reconstruction.Using super The major advantage of pixel is that computational efficiency is high.Compared with being represented using pixel, represented using super-pixel The image primitive quantity required for computing can be greatly reduced.

Superpixel segmentation method has many kinds, substantially can be divided into the method based on graph theory and be based under gradient The method of fall, illustrates to main method below:

The feature of normalized cuts method is the super-pixel controllable quantity producing, compact shape, each super Elemental area is substantially similar.Shortcoming is that algorithm is computationally intensive, speed is slow.

Mean shift method belongs to pattern search algorithm, and it passes through in pixel characteristic space iteration is mobile The heart, generates image segmentation.The super-pixel that mean shift method produces is accurate for localized variation.Shortcoming It is possible to produce irregular super-pixel it is impossible to directly control super-pixel quantity, size, compactness.

Quick shift dividing method is also a kind of pattern search dividing method, more a lot of soon than mean shift. This algorithm passes through continuous to promote each of pixel characteristic space data point, towards can make the gloomy density of handkerchief Estimate that the nearest pixel increasing moves, to realize the segmentation of image.Shortcoming is the super-pixel producing in shape It is not fixing on shape and quantity, and the compactness of super-pixel is also poor.

The basic thought of turbo pixel is to randomly select a quantity of seeds point on image and adopt water The method of flat collection is expanded, and controls the size of super-pixel block by limiting rate of increase.Shortcoming is point The super-pixel cut out is defined to homogeneous size, compactness and border degree of depending on, splitting speed extremely slow and It is relatively poor that border depends on situation.

There is a problem of in therefore existing super-pixel segmentation technology that splitting speed is slow, compactness is poor.

Content of the invention

The embodiment of the present invention provides a kind of superpixel segmentation method and device, is existed with overcoming in prior art The problem that splitting speed is slow, compactness is poor.

The embodiment of the present invention provides a kind of superpixel segmentation method, comprising:

According to the hyperspace numerical value of each pixel in n pending pixel and default closely Degree parameter, calculates the Euclidean distance between each two pixel, described compactness in described n pixel Parameter is the parameter for controlling super-pixel compactness；

According to the Euclidean distance between each two pixel in described n pixel and default the shortest Euclidean distance and central point probit, determine k initial center point in described n pixel, described K is less than n, described the shortest Euclidean distance be used for representing pixel to described pixel nearest in The shortest Euclidean distance of heart point；

Using annulus inequality, in described k initial center point, determine the nearest of other pixels respectively Central point, the pixel belonging to same central point is defined as same super-pixel, other pictures described Vegetarian refreshments is the pixel in described n pixel in addition to k initial center point.

The embodiment of the present invention also provides a kind of super-pixel segmentation device, comprising:

Computing module, for the hyperspace numerical value according to each pixel in n pending pixel And default tightness parameter, calculate in described n pixel the Euclidean between each two pixel away from From described tightness parameter is the parameter for controlling super-pixel compactness；

First determining module, for each two pixel in n pixel being calculated according to described computing module Euclidean distance between point, and default the shortest Euclidean distance and central point probit, at described n K initial center point is determined, described k is less than n, and described the shortest Euclidean distance is used for representing in pixel One pixel is to the shortest Euclidean distance of the nearest central point of described pixel；

Second determining module, for using annulus inequality, k determining in described first determining module Determine the nearest central point of other pixels in initial center point respectively, same central point will be belonged to Pixel is defined as same super-pixel, and other pixels described are except k in described n pixel Pixel outside initial center point.

Embodiment of the present invention superpixel segmentation method, n pixel is adopted hyperspace numerical value, and in advance If controlling the parameter of cluster compactness, calculate the Euclidean distance between each two pixel in n pixel, By default the shortest Euclidean distance and central point probit, select to determine k in described n pixel The initial center point of individual cluster, further, using annulus inequality, in described k initial center point The middle nearest central point determining other pixels respectively, will belong to the pixel of same central point afterwards It is defined as same super-pixel, realize the segmentation of k super-pixel.Because the embodiment of the present invention is using closely Degree parameter, can directly control the compactness of super-pixel, determine k due to employing the search of annulus inequality The central point of individual super-pixel, is greatly reduced the number of times of k means clustering method interior loop, accelerates super The splitting speed of pixel, solves the problems, such as in prior art that splitting speed is slow, compactness is poor.

Brief description

In order to be illustrated more clearly that the embodiment of the present invention or technical scheme of the prior art, below will be to reality The accompanying drawing applying required use in example or description of the prior art be briefly described it should be apparent that, under Accompanying drawing in the description of face is some embodiments of the present invention, for those of ordinary skill in the art, On the premise of not paying creative labor, other accompanying drawings can also be obtained according to these accompanying drawings.

Fig. 1 is the flow chart of superpixel segmentation method embodiment one of the present invention；

Fig. 2 is a kind of flow chart of the embodiment two of superpixel segmentation method of the present invention；

Fig. 3 is the schematic diagram of the annulus search adopting in the embodiment of the present invention；

Fig. 4 is a kind of structural representation of the embodiment three of present invention super-pixel segmentation device.

Specific embodiment

Purpose, technical scheme and advantage for making the embodiment of the present invention are clearer, below in conjunction with this Accompanying drawing in bright embodiment, is clearly and completely described to the technical scheme in the embodiment of the present invention, Obviously, described embodiment is a part of embodiment of the present invention, rather than whole embodiments.It is based on Embodiment in the present invention, those of ordinary skill in the art are obtained under the premise of not making creative work The every other embodiment obtaining, broadly falls into the scope of protection of the invention.

Fig. 1 is the flow chart of superpixel segmentation method embodiment one of the present invention, as shown in figure 1, this enforcement The method of example may include that

Step 101, according to the hyperspace numerical value of each pixel in n pending pixel and Default tightness parameter, calculates the Euclidean distance between each two pixel in described n pixel, Described tightness parameter is the parameter for controlling super-pixel compactness；

Wherein, described hyperspace numerical value includes 3-dimensional color value and 2 dimension space positional values；

Include when step 101 implements:

Using formulaCalculate the color between two pixels Apart from d_lab, wherein, i represents a pixel, and j represents one other pixel point, and l, a, b are pixel Value in lab color space；

Using formulaCalculate the locus distance between two pixels d_xy, wherein, x represents the position numerical value of pixel, and xy is pixel i, and j is on image x/y plane Position；

Using formulaCalculate Euclidean distance d between two pixels, wherein,Represent the interval between the central point of super-pixel.

Step 102, according to the Euclidean distance between each two pixel in described n pixel, and Default the shortest Euclidean distance and central point probit, during in described n pixel, determination k is initial Heart point, described k is less than n, and described the shortest Euclidean distance is used for representing a pixel to described pixel Nearest central point the shortest Euclidean distance.

The principle selecting initial center point in the embodiment of the present invention is between the initial center point choosing Distance is remote as far as possible, and that is, the Euclidean distance value between initial center point is big as far as possible.

Wherein, step 102 includes when implementing:

Step a, select any one pixel x in n pixel as the first initial center point, C1=x；

Step b, calculate other n-1 pixel respectively to the first initial center point c₁Euclidean distance d(c₁), wherein, other n-1 pixel be in n pixel in addition to the first initial center point x Other pixels；

Step c, according to step b calculate other n-1 pixel each to the first initial center point x Euclidean distance d (c₁), by d (c₁) higher value corresponding pixel x¹As the second initial center point c₂, c₂=x¹∈n；

Wherein, step c includes when implementing: first, calculates every in other n-1 pixel Individual pixel is to the first initial center point c₁Euclidean distance d (c₁) quadratic sum, i.e. σ d (c₁)², afterwards, A pixel x is selected in other n-1 pixel¹It is assumed that pixel x¹To the first initial center Point c₁Euclidean distance be d (x¹), by d (x¹) square, i.e. d (x¹)²Divided by σ d (c₁)²Alternatively Two initial center point c₂Probit, that is,Work as probitWhen bigger, corresponding picture Vegetarian refreshments x¹Elect the second initial center point c as₂Probability also bigger, conversely, working as probitLess When, corresponding pixel x¹Elect the second initial center point c as₂Probability also less；

Step d, calculate other n-2 pixel respectively to the second initial center point c₂Euclidean distance d(c₂), wherein, other n-2 pixel is except the first initial center point c in n pixel₁With Two initial center point c₂Outside other pixels；

Step e, according to step d calculate other n-2 pixel to the second initial center point c₂Europe Family name is apart from d (c₂), by d (c₂) it is higher value corresponding pixel x²As the 3rd initial center point c₃, c₃=x " ∈ n；Wherein, step e includes when implementing: first, calculates other n-2 pixel In each pixel to the second initial center point c₂Euclidean distance d (c₂) quadratic sum, i.e. σ d (c₂)², Afterwards, other n-2 pixel selects a pixel x²It is assumed that pixel x²Initial to second Central point c₂Euclidean distance be d (x²), by d (x²) square, i.e. d (x²)²Divided by σ d (c₂)²Make For selecting the 3rd initial center point c₃Probit, that is,Work as probitWhen bigger, Corresponding pixel x²Elect the 3rd initial center point c as₃Probability also bigger, conversely, working as probitMore hour, corresponding pixel x²Elect the 3rd initial center point c as₃Probability also less；

By that analogy, until selecting k-th initial center point, that is, select the general of k-th initial center point Rate value isWhen probit isWhen bigger, corresponding pixel x^k-1Elect kth as Individual initial center point c_kProbability also bigger, conversely, working as probitMore hour, corresponding picture Vegetarian refreshments x^k-1Elect k-th initial center point c as_kProbability also less；

Wherein, c_k-1For k-1 initial center point, d (c_k-1) it is except at the beginning of k-1 in n pixel Other pixels outside beginning central point, that is, each of n-k+1 pixel pixel is to kth -1 The Euclidean distance of initial center point；d(x^k-1) for the one of pixel x in n-k+1 pixel^k-1Arrive The Euclidean distance of -1 initial center point of kth.

Step 103, utilize annulus inequality, determine other pictures respectively in described k initial center point The nearest central point of vegetarian refreshments, by the pixel using same initial center point as its nearest central point Be defined as same super-pixel, other pixels described be in described n pixel except k initially in Pixel outside heart point.

Include when step 103 implements:

≤ | | x-c | |, wherein, x is for step a, described annulus inequality include the first inequality | | | x | |-| | c ' | | | Any pixel point in other pixels, c is set to distribute to the initial center point of pixel x, and described first Inequality represents with initial center point c for initial point one circle ring area of definition, if meeting described first Formula, then can determine that presence other central points c' than initial center point c closer to pixel x；

Step b, central point c' is set to distribute to initial center point c of pixel x in step a, that is, Say, make c'=c, judge whether again than c closer to pixel x central point c', if meeting described First inequality, then can determine that exist than c closer to pixel x central point c'；

Circulating repetition step b, until being unsatisfactory for described first inequality, or meets the second inequality | | | x-c | | ＜ | | | x | |-| | c ' | | | or | | x-c | |≤| | x-c ' | |, then can determine that and do not exist than c closer to pixel x Central point c', then described central point c is defined as the nearest central point of described pixel x, wherein, C ∈ k, c ' ∈ k；

It should be noted that step b is when implementing, it is that central point c' is replaced with the first of pixel x Beginning central point c is that is to say, that define a circle ring area with initial center point c' for initial point, if meeting not Equation | | | x | |-| | c " | |≤| | x-c'| |, then can determine that exist than c' closer to pixel x central point c "；Afterwards, by c " Replace with initial center point c of pixel x that is to say, that with initial center point c " define one for initial point Circle ring area, if meeting inequality | | | x | |-| | c " ' | |≤| | x-c " | |, can determine that presence than c " closer to pixel x Central point c " '；

By that analogy, until determining the immediate central point of pixel x, for example, work as satisfaction | | x-c " ' | | ＜ | | | x | |-| | c " " | | | or | | x-c " ' | |≤| | x-c " " | |, you can determine and do not exist than c " ' closer in pixel x Heart point c " ", then by central point c " ' determine the immediate central point of pixel x.

Afterwards, the pixel belonging to same central point is defined as same super-pixel.

Further, after step 103, that is, utilize annulus inequality, in described k initial center point The middle nearest central point determining other pixels respectively, the pixel belonging to same central point is determined After same super-pixel, also include:

Isolated pixel point is added the super-pixel closest with described isolated pixel point, wherein, described isolated Pixel refers to also not determine the pixel of nearest central point by above-mentioned steps 103；For example, picture is set Vegetarian refreshments x, to the longest Euclidean distance of nearest central point, calculates pixel x respectively to described k initial center The Euclidean distance of point, must beat k Euclidean distance it is assumed that this k Euclidean distance is all higher than arranging the longest Euclidean Distance, then can determine that pixel x is isolated pixel point, now can be added to this isolated pixel point and be somebody's turn to do The closest super-pixel of described isolated pixel point.

In another embodiment of the present invention, can also include after step 101:

Select k pixel as initial center point in described n pixel；

Each initial center point corresponding pixel cluster (quite with above-mentioned super-pixel), according to it His pixel arrives the Euclidean distance of described k initial center point respectively, and other pixels described are distributed to The initial center point corresponding pixel cluster nearest with other pixels described；

Redefine the central point of cluster according to default cluster average, wherein, described cluster average isn_iFor belonging to the pixel number in cluster i, p_xFor belonging to the data vector clustering i (such as 5 dimension space numerical value), i ∈ k；It should be noted that cluster here represents super-pixel；

So circulate, until object functionMeet end condition (such as Meet default threshold value), described n pixel is divided into the super-pixel of k cluster, wherein, x_jFor Data vector (as 5 dimension space numerical value), s_iFor x_jResiding cluster (super-pixel), μ_iFor clustering s_iMiddle picture The meansigma methodss of vegetarian refreshments.

N pixel is adopted hyperspace numerical value by the present embodiment first, and default control cluster is closely The parameter of degree, calculates the Euclidean distance between each two pixel in n pixel, by default Short Euclidean distance and central point probit, select to determine the initial of k cluster in described n pixel Central point, further, using annulus inequality, determines it in described k initial center point respectively The pixel belonging to same central point is defined as same by the nearest central point of his pixel afterwards Super-pixel, realizes the segmentation of k super-pixel.Because the embodiment of the present invention uses tightness parameter, permissible Directly control the compactness of super-pixel, further, the present embodiment also uses the search of annulus inequality and determines The central point of k super-pixel, can be greatly reduced the number of times of k means clustering method interior loop, greatly About reduce by 97% interior loop number of times, accelerate the splitting speed of super-pixel, prior art can be solved The problem that middle splitting speed is slow, compactness is poor.

Adopt specific embodiment below, the technical scheme of embodiment of the method shown in Fig. 1 is carried out specifically Bright.

Fig. 2 is a kind of flow chart of the embodiment two of superpixel segmentation method of the present invention, as shown in Fig. 2 The method of the present embodiment may include that

Step 201, view data is converted to cielab color space by primary color space, to picture The position xy of vegetarian refreshments is normalized conversion；

Step 202, by view data pixel formed labxy totally 5 dimension strong point；

The embodiment of the present invention carries out pixel cluster using in 5 dimension data space labxy.Wherein lab is figure As pixel is in the color vector of cielab color space, xy is position on image x/y plane for the image pixel Put.Because the ultimate range of two color-values is limited in cielab color space, and the sky of x/y plane Between distance related to picture size.So only after normalization xy space length, could be in 5 dimension data spaces Middle use Euclidean distance is compared.

Assume that setting super-pixel number is k, target is to divide an image into k approximately equalised super-pixel. For the image having n pixel, the approximate size of each super-pixel is n/k pixel, the center of super-pixel Interval between point is

Concrete distance calculating method:

$d_{lab}=\sqrt{{(i_j-i_i)}^2+{(a_j-a_i)}^2+{(b_j-b_i)}^2}---\left(1\right)$

$d_{xy}=\sqrt{{(x_j-x_i)}^2+{(x_j-x_i)}^2}---\left(2\right)$

$d=\sqrt{d_{lab}^2+{\left(\frac{d_{xy}}{s}\right)}^2{m}^2}---\left(3\right)$

Wherein d is the weighted sum of lab distance and normalized x/y plane distance.Parameter m can be used in controlling The compactness of super-pixel.

Step 203, using kmeans++ algorithms selection k means clustering algorithm initial center point； Wherein, k-means++ algorithm selects initial center according to the probability ratio of pixel to nearest central point Point, specifically includes:

(1) randomly choose initial center point c1=x from pixel n；

(2) setting d (x) is as the shortest Euclidean distance from a pixel x to nearest central point；

(3) select next central point c_i, wherein c_i=x ' ∈ x, its probit is

(4) 2 and 3 are repeated until selecting all of k center.

Step 204, using annulus search k means Method accelerate cluster；

The embodiment of the present invention annulus search k means Method be on the basis of k means Method improve and Come.

Wherein, k means Method is a kind of Unsupervised clustering algorithm, and its workflow is as follows:

(1) arbitrarily select k pixel from n pixel as initial cluster center, here initial Cluster centre is equivalent to the initial center point in embodiment illustrated in fig. 1；

(2) for other pixels, arrive the Euclidean distance of these cluster centres according to them, respectively by them Distribute to the cluster centre nearest with it, cluster centre here is equivalent to the center in embodiment illustrated in fig. 1 Point；

(3) according to the cluster centre that cluster mean value computation is new, cluster averagen_iFor belonging to The vector number of cluster i, p_xFor belonging to the vector of cluster i；

(4) so circulate, until target function value meets end condition, finally split data into k class.

In the present embodiment, using error sum of squares criterion function as object function:

$e=\arg \underset{s}{\min }{\σ}_{i=1}^k{\σ}_{x_j\&element;s_i}\left|\right|x_j-{\mathrm{\μ}}_i|{|}^2---\left(4\right)$

Wherein x_jFor data vector, s_iFor x_jResiding cluster, ν_iFor clustering s_iThe meansigma methodss at midpoint.

Due to there is the substantial amounts of step comparing with cluster centre, the speed of impact cluster in above-mentioned circulation. For this reason, the annulus search k means clustering method application annulus inequality of the embodiment of the present invention, for eliminating k The step that in means clustering method interior loop, majority is compared with cluster centre, Fig. 3 is in the embodiment of the present invention Using annulus search schematic diagram.

Interior loop relatively in only need to consider meet as lower inequality center c ':

|||x||-||c′|||≤||x-c|| (5)

Wherein x is data point, and c is the nearest central point distributing before.This inequality defines a circle Ring region, central point is initial point, including any than a closer to x central point c, a here represents center Point identification (as numbered).If formula 5 is unsatisfactory for,

| | x-c | | ＜ | | | x | |-| | c ' | | | (6)

≤||x-c′|| (7)

According to triangle inequality, the center c ' being unsatisfactory for formula 5 inequality will not be than c closer to x.I.e. x will Still it is assigned to cluster centre c, rather than be assigned to cluster centre c '.

The sign flag that annulus search k means clustering algorithm uses is described as follows:

C (j): cluster centre j (wherein 1≤j≤k)

S (j): cluster centre c (j) arrives its most paracentral distance

X (i): data point i (wherein 1≤i≤n)

A (i): mark data points i allocated center number

Apart from the upper bound between u (i): x (i) and c (a (i))

L (i): x (i) arrive it second paracentral apart from lower bound

Annulus search k means clustering algorithm flow process is as follows:

(1) | | x (i) | | is calculated to all i；

(2) circulation is until convergence

(21) for all of j, calculate s (j) and | | c (j) | |

(22) according to norm size to all centers ascending sort

(23) circulation i=1 arrives | x |

(231) make m=max (s (a (i))/2, l (i))

(232) if u (i)≤m jumps to next i

(233) make u (i)=d (x (i), c (a (i)))

(234) if u (i)≤m, jump to next i

(235) make l (i)=d (x (i), c (b (i)))

(236) make r=max (l (i), u (i))

(237) make j ← { j | | | | x (i) | |-| | c (j) | | |≤r }

(238) for all of j ∈ j, for cyclic search circle ring area

(2381) if d (x (i), c (j)) ＜ u (i), then find new nearest center

L (i)=u (i)；B (i)=a (i)

U (i)=d (x (i), c (j))；A (i)=j

(2382) otherwise, if d (x (i), c (j)) ＜ l (i), then in finding that new second is near

Heart l (i)=d (x (i), c (j)), a (i)=j

(2383) inequality judges to terminate

(239) loop ends

(24) loop ends

Each Dian Dao center mobile

Update the upper bound and lower bound

(3) loop ends

Search for k means Method using annulus and k means Method interior loop secondary can be greatly reduced Number, reduces by about 97% interior loop number of times, and lifting in efficiency is no less than 10 times.

The similar pixel generating after step 205, cluster is labeled as the same super-pixel after splitting；

Step 206, isolated pixel is added to neighbouring super-pixel.

View data is switched to cielab color space by primary color space by the present invention first, add into The normalized pixel position xy of row, forms labxy totally 5 dimension data space；For 5 dimension data spaces All data points, using the initial center point of kmeans++ algorithms selection k means clustering algorithm；Use Annulus search k means Method accelerates cluster；After the like numbers strong point generating after cluster is labeled as segmentation Same super-pixel；Isolated pixel is added neighbouring super-pixel.Compared with prior art, this In bright, super-pixel quantity is determined by the k value in k mean cluster, can directly control super-pixel quantity； Accelerating algorithm efficiency is searched for by annulus.

It should be noted that it is " initial using kmeans++ algorithms selection k means clustering algorithm in the present invention Central point " step is alternative step, uses this step this method instead and is still protected.In selecting initially The method of heart point has multiple, and kmeans++ algorithm is used herein.Other alternative initial methods Including, as seed point, random initializtion method, genetic algorithm is initial for the point selecting fixed-grid distance Change method etc..

Fig. 4 is a kind of structural representation of the embodiment three of present invention super-pixel segmentation device, as Fig. 4 institute Show, the device of the present embodiment may include that

Computing module 41, for the hyperspace number according to each pixel in n pending pixel Value and default tightness parameter, calculate the Euclidean between each two pixel in described n pixel Distance, described tightness parameter is the parameter for controlling super-pixel compactness；

First determining module 42, for each two picture in n pixel being calculated according to described computing module Euclidean distance between vegetarian refreshments, and default the shortest Euclidean distance and central point probit, in described n K initial center point is determined, described k is less than n, and described the shortest Euclidean distance is used for table in individual pixel Show a pixel to the shortest Euclidean distance of the nearest central point of described pixel；

Second determining module 43, for using annulus inequality, the k determining in described first determining module Determine the nearest central point of other pixels in individual initial center point respectively, same central point will be belonged to Pixel be defined as same super-pixel, other pixels described be described n pixel in except k Pixel outside individual initial center point.

For example, above-mentioned hyperspace numerical value includes 3-dimensional color value and 2 dimension space positional values；

Wherein, computing module 41 specifically can be used for:

Using formulaCalculate the color between two pixels Distance, wherein, i represents a pixel, and j represents one other pixel point, and l, a, b exist for pixel Value in lab color space；

Using formulaCalculate the locus distance between two pixels, Wherein, x represents the position numerical value of pixel, and xy is pixel i, position on image x/y plane for the j；

Using formulaCalculate the Euclidean distance between two pixels, wherein,Represent the interval between the central point of super-pixel.

For example, described first determining module 42 specifically can be used for:

Step a, select any one pixel x in n pixel as the first initial center point, C1=x；

Step b, according to each pixel in other n-1 pixel to the first initial center point c₁Europe Family name is apart from d (c₁), calculate each pixel in other n-1 pixel to the first initial center point c₁'s Euclidean distance d (c₁) quadratic sum, i.e. σ d (c₁)², wherein, other n-1 pixel is n pixel Other pixels in addition to the first initial center point x in point；

Step c, select one of pixel x in other n-1 pixel¹If, pixel x¹Arrive First initial center point c₁Euclidean distance be d (x¹), by d (x¹) square, i.e. d (x¹)²Divided by σ d (c₁)² Alternatively the second initial center point c₂Probit, that is,Work as probitWhen bigger, Corresponding pixel x¹Elect the second initial center point c as₂Probability also bigger, conversely, working as probitMore hour, corresponding pixel x¹Elect the second initial center point c as₂Probability also less；

Step d, according to each pixel in other n-2 pixel to the second initial center point c₂'s Euclidean distance d (c₂), calculate each pixel in other n-2 pixel to the second initial center point c₂ Euclidean distance d (c₂) quadratic sum, i.e. σ d (c₂)², wherein, other n-2 pixel is n picture Except the first initial center point c in vegetarian refreshments₁With the second initial center point c₂Outside other pixels；

Step e, in other n-2 pixel select a pixel x²If, pixel x²To second Initial center point c₂Euclidean distance be d (x²), by d (x²) square, i.e. d (x²)²Divided by σ d (c₂)² Alternatively the 3rd initial center point c₃Probit, that is,Work as probitWhen bigger, Corresponding pixel x²Elect the 3rd initial center point c as₃Probability bigger, conversely, working as probit More hour, corresponding pixel x²Elect the 3rd initial center point c as₃Probability less；

By that analogy, until selecting k-th initial center point, that is, select the general of k-th initial center point Rate value isWhen probit isWhen bigger, corresponding pixel x^k-1Elect kth as Individual initial center point c_kProbability also bigger, conversely, working as probitMore hour, corresponding picture Vegetarian refreshments x^k-1Elect k-th initial center point c as_kProbability also less；

Wherein, c_k-1For k-1 initial center point, d (c_k-1) it is except at the beginning of k-1 in n pixel Other pixels outside beginning central point, that is, each of n-k+1 pixel pixel is to kth -1 The Euclidean distance of initial center point；d(x^k-1) for the one of pixel x in n-k+1 pixel^k-1Arrive The Euclidean distance of -1 initial center point of kth.

For example, described second determining module 43 specifically can be used for:

≤ | | x-c | |, wherein, x is for step a, described annulus inequality include the first inequality | | | x | |-| | c ' | | | Any pixel point in other pixels, c is set to distribute to the initial center point of pixel x, and described first Inequality represents with initial center point c for initial point one circle ring area of definition, if meeting described first Formula, then can determine that exist than c closer to pixel x central point c'；

Step b, central point c' is set to distribute to initial center point c of pixel x, even c'=c, if Meet described first inequality, then can determine that exist than c closer to pixel x central point c'；

Circulating repetition step b, until being unsatisfactory for described first inequality, or meets the second inequality | | x-c | | ＜ | | | x | |-| | c ' | | | or | | x-c | |≤x-c ' | |, then can determine that and do not exist than c closer to pixel x Central point c', then described central point c is defined as the nearest central point of described pixel x, wherein, C ∈ k, c' ∈ k.Step b implement the associated description that may be referred in embodiment illustrated in fig. 1, no longer Repeat.

For example, described device also includes: the 3rd determining module 44, if for other pixels described Also there is the pixel not determining nearest central point in point, then do not determined the pixel of nearest central point Point as isolated pixel point, and by described isolated pixel point add closest with described isolated pixel point by The super-pixel that second determining module determines.

The device of the present embodiment, can be used for executing the technical scheme of embodiment of the method shown in Fig. 1 or Fig. 2, It is realized, and principle is similar with technique effect, and here is omitted.

Described above illustrate and describes some preferred embodiments of the present utility model, but as it was previously stated, should When understanding that this utility model is not limited to form disclosed herein, it is not to be taken as to other embodiment Exclusion, and can be used for various other combinations, modification and environment, and can be in utility model described herein In contemplated scope, it is modified by the technology or knowledge of above-mentioned teaching or association area.And people from this area The change that carried out of member and change, then all should in this practicality newly without departing from spirit and scope of the present utility model In the protection domain of type claims.

Claims (10)

1. a kind of superpixel segmentation method it is characterised in that:

According to the hyperspace numerical value of each pixel in n pending pixel and default closely Degree parameter, calculates the Euclidean distance between each two pixel, described compactness in described n pixel Parameter is the parameter for controlling super-pixel compactness；

According to the Euclidean distance between each two pixel in described n pixel and default the shortest Euclidean distance and central point probit, determine k initial center point in described n pixel, described K is less than n, described the shortest Euclidean distance be used for representing pixel to described pixel nearest in The shortest Euclidean distance of heart point；

Using annulus inequality, in described k initial center point, determine the nearest of other pixels respectively Central point, the pixel belonging to same central point is defined as same super-pixel, other pictures described Vegetarian refreshments is the pixel in described n pixel in addition to k initial center point.

2. method according to claim 1 is it is characterised in that described hyperspace numerical value includes 5 Dimension space numerical value, specifically includes 3-dimensional color value and 2 dimension space positional values；

In the described n pixel of described calculating, the Euclidean distance between each two pixel includes:

Using formula $d_{lab}=\sqrt{{(l_j-l_i)}^2+{(a_j-a_i)}^2+{(b_j-b_i)}^2},$ Calculate the color between two pixels Distance, wherein, i represents a pixel, and j represents one other pixel point, and l, a, b exist for pixel Value in lab color space；

Using formulaCalculate the locus distance between two pixels, Wherein, x represents the position numerical value of pixel, and xy is pixel i, position on image x/y plane for the j；

Using formulaCalculate the Euclidean distance between two pixels, wherein,Represent the interval between the central point of super-pixel.

3. method according to claim 1 is it is characterised in that according in described n pixel Euclidean distance between each two pixel, and default the shortest Euclidean distance and central point probit, K initial center point is determined in described n pixel, comprising:

Step a, select any one pixel x in n pixel as the first initial center point, C1=x；

Step b, according to each pixel in other n-1 pixel to the first initial center point c₁Europe Family name is apart from d (c₁), calculate each pixel in other n-1 pixel to the first initial center point c₁'s Euclidean distance d (c₁) quadratic sum, i.e. σ d (c₁)², wherein, other n-1 pixel is n pixel Other pixels in addition to the first initial center point x in point；

Step c, select one of pixel x in other n-1 pixel¹If, pixel x¹Arrive First initial center point c₁Euclidean distance be d (x¹), by d (x¹) square, i.e. d (x¹)²Divided by σ d (c₁)² Alternatively the second initial center point c₂Probit, that is,Work as probitWhen bigger, Pixel x¹Elect the second initial center point c as₂Probability also bigger；

Step d, according to each pixel in other n-2 pixel to the second initial center point c₂'s Euclidean distance d (c₂), calculate each pixel in other n-2 pixel to the second initial center point c₂ Euclidean distance d (c₂) quadratic sum, i.e. σ d (c₂)², wherein, other n-2 pixel is n picture Except the first initial center point c in vegetarian refreshments₁With the second initial center point c₂Outside other pixels；

Step e, in other n-2 pixel select a pixel x²If, pixel x²To second Initial center point c₂Euclidean distance be d (x²), by d (x²) square, i.e. d (x²)²Divided by σ d (c₂)² Alternatively the 3rd initial center point c₃Probit, that is,Work as probitWhen bigger, Pixel x²Elect the 3rd initial center point c as₃Probability also bigger；

By that analogy, until selecting k-th initial center point, that is, select the general of k-th initial center point Rate value isWherein, c_k-1For k-1 initial center point, d (c_k-1) in n pixel Other pixels in addition to k-1 initial center point, i.e. each of n-k+1 pixel pixel Point is to the Euclidean distance of -1 initial center point of kth；d(x^k-1) for wherein in n-k+1 pixel Individual pixel x^k-1Euclidean distance to -1 initial center point of kth.

4. the method according to any one of claim 1-3 it is characterised in that utilize annulus inequality, The nearest central point of other pixels is determined respectively in described k initial center point, comprising:

≤ | | x-c | |, wherein, x is for step a, described annulus inequality include the first inequality | | | x | |-| | c ' | | | Any pixel point in other pixels, c is set to distribute to the initial center point of pixel x, and described first Inequality represents with initial center point c for initial point one circle ring area of definition, if meeting described first Formula, then can determine that exist than c closer to pixel x central point c'；

Step b, central point c' is set to distribute to initial center point c of pixel x, even c'=c, if full Described first inequality of foot, then can determine that exist than c closer to pixel x central point c'；

Circulating repetition step b, until being unsatisfactory for described first inequality, or meets the second inequality | | x-c | | ＜ | | | x | |-| | c ' | | | or | | x-c | |≤| | x-c ' | |, then can determine that and do not exist than c closer to pixel x Central point c', then described central point c is defined as the nearest central point of described pixel x, wherein, C ∈ k, c ' ∈ k.

5. method according to claim 1 is it is characterised in that utilize annulus inequality, described Determine the nearest central point of other pixels in k initial center point respectively, same center will be belonged to After the pixel of point is defined as same super-pixel, also include:

If also there is, in other pixels described, the pixel not determining nearest central point, do not have described There is the pixel determining nearest central point as isolated pixel point, and described isolated pixel point is added and institute State the closest super-pixel of isolated pixel point.

6. a kind of super-pixel segmentation device is it is characterised in that include:

Computing module, for the hyperspace numerical value according to each pixel in n pending pixel And default tightness parameter, calculate in described n pixel the Euclidean between each two pixel away from From described tightness parameter is the parameter for controlling super-pixel compactness；

First determining module, for each two pixel in n pixel being calculated according to described computing module Euclidean distance between point, and default the shortest Euclidean distance and central point probit, at described n K initial center point is determined, described k is less than n, and described the shortest Euclidean distance is used for representing in pixel One pixel is to the shortest Euclidean distance of the nearest central point of described pixel；

Second determining module, for using annulus inequality, k determining in described first determining module Determine the nearest central point of other pixels in initial center point respectively, same central point will be belonged to Pixel is defined as same super-pixel, and other pixels described are except k in described n pixel Pixel outside initial center point.

7. device according to claim 6 is it is characterised in that described hyperspace numerical value includes 5 Dimension space numerical value, specifically includes 3-dimensional color value and 2 dimension space positional values；

Described computing module specifically for:

Using formula $d_{lab}=\sqrt{{(l_j-l_i)}^2+{(a_j-a_i)}^2+{(b_j-b_i)}^2},$ Calculate the color between two pixels Distance, wherein, i represents a pixel, and j represents one other pixel point, and l, a, b exist for pixel Value in lab color space；

Using formulaCalculate the locus distance between two pixels, Wherein, x represents the position numerical value of pixel, and xy is pixel i, position on image x/y plane for the j；

Using formulaCalculate the Euclidean distance between two pixels, wherein,Represent the interval between the central point of super-pixel.

8. device according to claim 6 is it is characterised in that described first determining module is specifically used In:

Step a, select any one pixel x in n pixel as the first initial center point, C1=x；

Step b, according to each pixel in other n-1 pixel to the first initial center point c₁Europe Family name is apart from d (c₁), calculate each pixel in other n-1 pixel to the first initial center point c₁'s Euclidean distance d (c₁) quadratic sum, i.e. σ d (c₁)², wherein, other n-1 pixel is n pixel Other pixels in addition to the first initial center point x in point；

Step c, select one of pixel x in other n-1 pixel¹If, pixel x¹Arrive First initial center point c₁Euclidean distance be d (x¹), by d (x¹) square, i.e. d (x¹)²Divided by σ d (c₁)² Alternatively the second initial center point c₂Probit, that is,Work as probitWhen bigger, Pixel x¹Elect the second initial center point c as₂Probability also bigger；

Step d, according to each pixel in other n-2 pixel to the second initial center point c₂'s Euclidean distance d (c₂), calculate each pixel in other n-2 pixel to the second initial center point c₂ Euclidean distance d (c₂) quadratic sum, i.e. σ d (c₂)², wherein, other n-2 pixel is n picture Except the first initial center point c in vegetarian refreshments₁With the second initial center point c₂Outside other pixels；

Step e, in other n-2 pixel select a pixel x²If, pixel x²To second Initial center point c₂Euclidean distance be d (x²), by d (x²) square, i.e. d (x²)²Divided by σ d (c₂)² Alternatively the 3rd initial center point c₃Probit, that is,Work as probitWhen bigger, Pixel x²Elect the 3rd initial center point c as₃Probability also bigger；

By that analogy, until selecting k-th initial center point, that is, select the general of k-th initial center point Rate value isWherein, c_k-1For k-1 initial center point, d (c_k-1) in n pixel Other pixels in addition to k-1 initial center point, i.e. each of n-k+1 pixel pixel Point is to the Euclidean distance of -1 initial center point of kth；d(x^k-1) for wherein in n-k+1 pixel Individual pixel x^k-1Euclidean distance to -1 initial center point of kth.

9. the device according to any one of claim 6-8 is it is characterised in that described second determines mould Block specifically for:

≤ | | x-c | |, wherein, x is for step a, described annulus inequality include the first inequality | | | x | |-| | c ' | | | Any pixel point in other pixels, c is set to distribute to the initial center point of pixel x, and described first Inequality represents with initial center point c for initial point one circle ring area of definition, if meeting described first Formula, then can determine that exist than c closer to pixel x central point c'；

Step b, central point c' is set to distribute to initial center point c of pixel x, even c'=c, if full Described first inequality of foot, then can determine that exist than c closer to pixel x central point c'；

Circulating repetition step b, until being unsatisfactory for described first inequality, or meets the second inequality | | x-c | | ＜ | | | x | |-| | c ' | | | or | | x-c | |≤| | x-c ' | |, then can determine that and do not exist than c closer to pixel x Central point c', then described central point c is defined as the nearest central point of described pixel x, wherein, C ∈ k, c ' ∈ k.

10. device according to claim 6 is it is characterised in that also include:

3rd determining module, if do not determine nearest central point for also existing in other pixels described Pixel, then using the described pixel not determining nearest central point as isolated pixel point, and will be described Isolated pixel point adds the super-pixel closest with described isolated pixel point.