CN105931244A - Supervision-free image matting method and apparatus - Google Patents

Abstract

The invention discloses a supervision-free image matting method and apparatus. Through performing super pixel segmentation on an image to be matted, the image to be matted is segmented into multiple super pixel areas, based on detected positions of angular points in an angular point set of the image to be matted, a convex hull area corresponding to the angular point set is determined, accordingly, through outward extension of the convex hull area, a prior super pixel area is determined, furthermore, based on a position relation between the prior super pixel area and a non-prior super pixel area, an unknown super pixel area is determined, accordingly, based on the determined prior super pixel area, the non-prior super pixel area and the unknown super pixel area, a constraint gray scale image is generated, and based on the generated constraint gray scale image, image matting processing is performed on the image to be matted. By using the scheme provided by the invention, the accuracy of automatically obtaining constrain information from the image to be matted is improved.

Description

A kind of without the stingy drawing method of supervision and device

Technical field

The present invention relates to technical field of computer vision, at field of artificial intelligence, and image Reason technical field, particularly relates to a kind of without the stingy drawing method of supervision and device.

Background technology

Scratching figure and extract the technology in foreground target region from image or video, it is computer vision One of key problem with field of artificial intelligence research.Owing to stingy diagram technology is at image procossing and video Important application in terms of editor, it is also that the focus of current technical field of image processing research is asked simultaneously One of topic.

Currently, according to whether use man-machine interactively, stingy drawing method can be divided into two big classes: semi-supervised Scratch drawing method, scratch drawing method without supervision.Semi-supervised stingy drawing method is by manually providing constraints Complete the extraction of mask, be a focus of current research, achieve much significantly one-tenth in recent years Really.Unsupervised stingy drawing method automatically extracts constraint information by machine and supports the generation of mask.

At present, constraint information automatically extract be research difficult point, be without supervision scratch drawing method development Bottleneck place.Therefore, very limited for the research scratching nomography without supervision.But, due to Its huge potential using value is starting the most studied persons and is being paid close attention to.But, how Automatically can obtain constraint information, accurately from image, and then the picture generation of covering in auxiliary later stage is One challenging problem.

The one being currently known is scratched drawing method and is scratched nomography for spectrum, by image correspondence Laplce's square The minimal characteristic vector of battle array calculates basis T_BlurMasked parts, and application shade component construction semantically has The prospect of meaning covers picture.But, obtain the most in actual applications based on semantic subassembly selection Mistake cover picture, tracing it to its cause is that constraint information is not accurate enough.

In sum, the stingy drawing method being currently known exists and obtains constraint information (TriMap) standard The problem that really property is relatively low.

Summary of the invention

The embodiment of the present invention provides a kind of and scratches drawing method and device without supervision, presently, there are in order to solving The relatively low problem of constraint information accuracy that obtains during FIG pull handle.

The embodiment of the present invention provides a kind of and scratches drawing method without supervision, including:

Treat stingy figure image and carry out super-pixel segmentation, obtain multiple super-pixel region；

Described figure image to be scratched is carried out Corner Detection, obtain described in angle point collection in figure image to be scratched Close；

Based on the position of angle point in described angle point set, determine the convex closure district that described angle point set is corresponding Territory；

In the plurality of super-pixel region, determine the super-pixel region of described convex closure region overlay to The most adjacent super-pixel region, the super-pixel region of described convex closure region overlay is the most adjacent with described Super-pixel region, as priori super-pixel region；

For each priori super-pixel region, if the neighbouring super pixels district in this priori super-pixel region Territory exists non-priori super-pixel region, then this priori super-pixel region is defined as unknown super-pixel Region；

For each non-priori super-pixel region, if the adjacent super picture in this non-priori super-pixel region Element region exists unknown super-pixel region, then this non-priori super-pixel region is defined as unknown super Pixel region；

Priori super-pixel region, non-priori super-pixel region and unknown super-pixel determined by based on Region, generates constraint gray level image；

Based on described constraint gray level image, described figure image to be scratched is carried out FIG pull handle.

Further, based on determined by priori super-pixel region, non-priori super-pixel region and Unknown super-pixel region, generates constraint gray level image, including:

The pixel value of the pixel in priori super-pixel region determined by by is defined as representing white Pixel value, by determined by the pixel value of pixel in non-priori super-pixel region be defined as representing black The pixel value of look, by determined by the pixel value of pixel in unknown super-pixel region be defined as representing Black and the pixel value of white Intermediate grey, obtain retraining gray level image.

The embodiment of the present invention also provides for a kind of without supervising stingy map device, including:

First segmentation module, is used for treating stingy figure image and carries out super-pixel segmentation, obtain multiple super picture Element region；

First Corner Detection module, for described figure image to be scratched is carried out Corner Detection, obtains institute State the angle point set in figure image to be scratched；

First convex closure area determination module, is used for based on the position of angle point in described angle point set, really The convex closure region that fixed described angle point set is corresponding；

First expansion module, in the plurality of super-pixel region, determines described convex closure region The most adjacent super-pixel region, super-pixel region covered, the super-pixel of described convex closure region overlay Region and the most adjacent described super-pixel region, as priori super-pixel region；

First area determines module, for for each priori super-pixel region, if this priori surpasses The neighbouring super pixels region of pixel region exists non-priori super-pixel region, then this priori is surpassed picture Element region is defined as unknown super-pixel region；

Second area determines module, is used for for each non-priori super-pixel region, if this non-elder generation Test and the neighbouring super pixels region in super-pixel region exists unknown super-pixel region, then by this non-priori Super-pixel region is defined as unknown super-pixel region；

First constraint information generation module, for based on determined by priori super-pixel region, non-elder generation Test super-pixel region and unknown super-pixel region, generate constraint gray level image；

First FIG pull handle module, for based on described constraint gray level image, to described figure figure to be scratched As carrying out FIG pull handle.

In the said method that the embodiment of the present invention provides, carry out super-pixel divide by treating stingy figure image Cut, figure image to be scratched is divided into multiple super-pixel region, and based on the figure figure to be scratched detected The position of angle point in the angle point set of picture, determines the convex closure region that angle point set is corresponding, and then leads to Cross convex closure region outward expansion, determine priori super-pixel region, and be based further on priori Position relationship between super-pixel region and non-priori super-pixel region, determines unknown super-pixel district Territory, thus realize based on determined by priori super-pixel region, non-priori super-pixel region and not Know super-pixel region, generate constraint gray level image, and based on the constraint gray level image generated, from treating Scratch and figure image extracts foreground target image.During generating constraint gray level image, introduce Super-pixel region, and the process of angle point convex closure detection such that it is able to determine table more accurately Show the priori super-pixel region of foreground image, and then improve automatic from figure image to be scratched acquisition about The accuracy of bundle information.

The embodiment of the present invention also provides for a kind of without supervising stingy drawing method, including:

Treat stingy figure image and carry out super-pixel segmentation, obtain multiple super-pixel region；

Characteristic vector based on the plurality of super-pixel region, determines in the plurality of super-pixel region Characteristic relation between each two super-pixel region；

Described figure image to be scratched is carried out Corner Detection, obtain described in angle point collection in figure image to be scratched Close；

Based on the position of angle point in described angle point set, determine the convex closure district that described angle point set is corresponding Territory；

In the plurality of super-pixel region, determine the super-pixel region of described convex closure region overlay to The most adjacent super-pixel region, the super-pixel region of described convex closure region overlay is the most adjacent with described Super-pixel region, as priori super-pixel region, the wherein priori label in priori super-pixel region Value is 1, and the priori label value in non-priori super-pixel region is 0；

Utilize condition random field theory, priori label value based on the plurality of super-pixel region, with And characteristic relation between each two super-pixel region in the plurality of super-pixel region, to described many The priori label value in individual super-pixel region is optimized, and the priori label value after optimization is the super picture of 1 Element region is as optimizing priori super-pixel region；

For each optimization priori super-pixel region, if this optimization priori super-pixel region is adjacent Super-pixel region exists unoptimizable priori super-pixel region, then by this optimization priori super-pixel region It is defined as unknown super-pixel region；

For each unoptimizable priori super-pixel region, if this unoptimizable priori super-pixel region Neighbouring super pixels region exists unknown super-pixel region, then by this unoptimizable priori super-pixel region It is defined as unknown super-pixel region；

Priori super-pixel region, unoptimizable priori super-pixel region and not is optimized determined by based on Know super-pixel region, generate constraint gray level image；

Based on described constraint gray level image, described figure image to be scratched is carried out FIG pull handle.

Further, characteristic vector based on the plurality of super-pixel region, determine the plurality of super Characteristic relation between each two super-pixel region in pixel region, including:

Equation below is used to determine in the plurality of super-pixel region between each two super-pixel region Characteristic relation:

$E_{ij}=\left\{\begin{array}{cc}\exp (-W(i,j)\left|\right|v_i-v_j\left|{|}^2\right)& (v_j\∈C_i)\\ 0& \left(v_j\∉C_i\right)\end{array}\right.;$

Wherein, E_ijRepresent that in the plurality of super-pixel region, i-th super-pixel region surpasses with jth Characteristic relation value between pixel region, v_iRepresent the characteristic vector in i-th super-pixel region, v_jTable Show the characteristic vector in jth super-pixel region, C_iRepresenting has predetermined with i-th super-pixel region The super-pixel regional ensemble of justice annexation；

W (i, j)=(| | v_i-v_k(i)|g||v_k(j)-v_j||)^-1；

Wherein, n is C_iThe quantity in the super-pixel region included, v_kI () is for meet following relational expression C_iIn the characteristic vector in a super-pixel region；

$v_k\left(i\right)=\arg \min \underset{l=1}{\overset{n}{\Σ}}\left|\right|v_k\left(i\right)-v_l\left(i\right){|}_2^2.$

Further, characteristic vector v in super-pixel region is expressed as v=[L, a, b, R, G, B, x, y], its In, L, a, b represent super-pixel region mean value of each passage under Lab color space respectively, R, G, B represent super-pixel region mean value of each passage under RGB color respectively, x, Y represents the center in super-pixel region abscissa in described figure image to be scratched and vertical seat respectively Mark.

Further, utilize condition random field theory, priori based on the plurality of super-pixel region Characteristic relation between each two super-pixel region in label value, and the plurality of super-pixel region, The priori label value in the plurality of super-pixel region is optimized, including:

For energy equationSeek optimal solution, obtain optimum SolveWherein, k_iRepresent the weight that i-th super-pixel region is corresponding, l_iRepresent The priori label value in i-th super-pixel region, E_ijRepresent i-th in the plurality of super-pixel region Characteristic relation value between super-pixel region and jth super-pixel region, λ is balance factor,Table Showing the value of optimal solution corresponding to i-th super-pixel region, N is the number in the plurality of super-pixel region Amount；

Use self adaptation two value-based algorithm, to described optimal solution y^*Process, obtain the plurality of super Priori label value Y=[y after the optimization of pixel region₁,y₂,…,y_N], wherein, y_iRepresent that i-th surpasses Priori label value after the optimization of pixel region.

Further, for energy equationAsk optimum Solve, obtain following optimal solution:

y^*=(K+ λ P)^-1(λKL)；

Wherein, P be described in the Laplacian Matrix of figure image to be scratched, K is unit matrix, and L is described The vector of the priori label value composition in multiple super-pixel regions；

P=D-E；

Wherein, D=diag{d₁₁,d₂₂,…,d_NN}；

$d_{ii}=\underset{j}{\Σ}{E}_{ij}.$

Further, based on determined by optimize priori super-pixel region, unoptimizable priori super-pixel Region and unknown super-pixel region, generate constraint gray level image, including:

The pixel value optimizing the pixel in priori super-pixel region determined by by is defined as representing white The pixel value of look, by determined by the pixel value of pixel in unoptimizable priori super-pixel region determine For representing the pixel value of black, by determined by the pixel value of pixel in unknown super-pixel region true It is set to the pixel value representing black with white Intermediate grey, obtains retraining gray level image.

The embodiment of the present invention also provides for a kind of without supervising stingy map device, including:

Second segmentation module, is used for treating stingy figure image and carries out super-pixel segmentation, obtain multiple super picture Element region；

Characteristic relation determines module, for characteristic vector based on the plurality of super-pixel region, really Determine the characteristic relation between each two super-pixel region in the plurality of super-pixel region；

Second Corner Detection module, for described figure image to be scratched is carried out Corner Detection, obtains institute State the angle point set in figure image to be scratched；

Second convex closure area determination module, is used for based on the position of angle point in described angle point set, really The convex closure region that fixed described angle point set is corresponding；

Second expansion module, in the plurality of super-pixel region, determines described convex closure region The most adjacent super-pixel region, super-pixel region covered, the super-pixel of described convex closure region overlay Region and the most adjacent described super-pixel region, as priori super-pixel region, wherein priori surpasses The priori label value of pixel region is 1, and the priori label value in non-priori super-pixel region is 0；

Optimize module, be used for utilizing condition random field theory, based on the plurality of super-pixel region Feature between each two super-pixel region in priori label value, and the plurality of super-pixel region Relation, is optimized the priori label value in the plurality of super-pixel region, the first standard inspection after optimization Label value is that the super-pixel region of 1 is as optimizing priori super-pixel region；

3rd area determination module, for for each optimization priori super-pixel region, if this is excellent Change in the neighbouring super pixels region in priori super-pixel region and there is unoptimizable priori super-pixel region, then This optimization priori super-pixel region is defined as unknown super-pixel region；

4th area determination module, for for each unoptimizable priori super-pixel region, if should The neighbouring super pixels region in unoptimizable priori super-pixel region exists unknown super-pixel region, then will This unoptimizable priori super-pixel region is defined as unknown super-pixel region；

Second constraint information generation module, for based on determined by optimize priori super-pixel region, Unoptimizable priori super-pixel region and unknown super-pixel region, generate constraint gray level image；

Second FIG pull handle module, for based on described constraint gray level image, to described figure figure to be scratched As carrying out FIG pull handle.

In the said method that the embodiment of the present invention provides, carry out super-pixel divide by treating stingy figure image Cut, figure image to be scratched is divided into multiple super-pixel region, and determines each two super-pixel region Between characteristic relation, and based on the position of angle point in the angle point set of the figure image to be scratched detected Put, determine the convex closure region that angle point set is corresponding, and then again by convex closure region outward expansion, Determine priori super-pixel region, then utilize condition random field theory, to the priori determined The priori label value in super-pixel region is optimized, in order to determine the optimization more tallied with the actual situation Priori super-pixel region, and be based further on optimizing priori super-pixel region and unoptimizable priori and surpass Position relationship between pixel region, determines unknown super-pixel region, thus realize based on really Fixed optimization priori super-pixel region, unoptimizable priori super-pixel region and unknown super-pixel region, Generate constraint gray level image, and based on the constraint gray level image generated, extract from figure image to be scratched Foreground target image.During generating constraint gray level image, introduce super-pixel region, Position relationship between super-pixel region, angle point convex closure detects, and utilizes condition random field theory Optimize the process of the priori label value in the priori super-pixel region initially determined that such that it is able to more accurate Determine the priori super-pixel region representing foreground image, and then improve from figure image to be scratched Automatically the accuracy of constraint information is obtained.

Other features and advantage will illustrate in the following description, and, partly Become apparent from specification, or understand by implementing the application.The purpose of the application Can be by being referred in particular in the specification write, claims and accompanying drawing with other advantages The structure gone out realizes and obtains.

Accompanying drawing explanation

Accompanying drawing is for providing a further understanding of the present invention, and constitutes a part for specification, It is used for explaining the present invention together with the embodiment of the present invention, is not intended that limitation of the present invention.Attached In figure:

The flow chart scratching drawing method without supervision that Fig. 1 provides for the embodiment of the present invention 1；

The flow chart scratching drawing method without supervision that Fig. 2 provides for the embodiment of the present invention 2；

Fig. 3 in the embodiment of the present invention 3 provide based on constraint gray level image, from figure image to be scratched The flow chart of middle extraction foreground target image；

The structural representation scratching map device without supervision that Fig. 4 provides for the embodiment of the present invention 4；

The structural representation scratching map device without supervision that Fig. 5 provides for the embodiment of the present invention 5.

Detailed description of the invention

The realization side of acquisition constraint information accuracy during FIG pull handle is improved in order to be given Case, embodiments provides and scratches drawing method and device without supervision, below in conjunction with Figure of description The preferred embodiments of the present invention are illustrated, it will be appreciated that preferred embodiment described herein It is merely to illustrate and explains the present invention, being not intended to limit the present invention.And in situation about not conflicting Under, the embodiment in the application and the feature in embodiment can be mutually combined.

Embodiment 1:

The embodiment of the present invention 1 provides a kind of and scratches drawing method without supervision, as it is shown in figure 1, can include Following steps:

Step 11, treat stingy figure image and carry out super-pixel segmentation, obtain multiple super-pixel region.

In this step, the various super-pixel partitioning algorithms being currently known can be used to treat stingy figure image Carry out super-pixel segmentation, such as, use linear iteraction clustering algorithm (SLIC) to treat stingy figure image Carry out super-pixel segmentation, obtain multiple super-pixel region.

Step 12, treat stingy figure image and carry out Corner Detection, obtain the angle point collection in figure image to be scratched Close.

Strict sequencing is not had between this step 12 and above-mentioned steps 11.

Step 13, based on the position of angle point in angle point set, determine the convex closure district that angle point set is corresponding Territory.

This step is equivalent to be attached angle point peripherally located for the position in angle point set, and even The curve connect expands outwardly, obtained convex closure region, compares straight line and connects the region model drawn a circle to approve Enclose bigger, so that initial foreground area scope determined by follow-up is the biggest, in order to reduce The final inaccuracy obtaining constraint information.

Step 14, in multiple super-pixel regions, determine the super-pixel region of convex closure region overlay to The most adjacent super-pixel region, i.e. convex closure region outward expansion one ring, the super picture of convex closure region overlay Element region and the most adjacent super-pixel region, as priori super-pixel region.

Understand based on priori, the boundary point of the foreground target image in a sub-picture, be mostly Angle point, so, in the embodiment of the present invention 1, based on priori, by above-mentioned steps 12-14 In Corner Detection etc. process step, the region that the priori super-pixel region determined is covered, Initial foreground area can be expressed as.

Step 15, for each priori super-pixel region, if this priori super-pixel region is adjacent Super-pixel region exists non-priori super-pixel region, then this priori super-pixel region is defined as not Know super-pixel region.

Step 16, by above-mentioned steps 15, carry out processing it for each priori super-pixel region After, in this step, for each non-priori super-pixel region, if this non-priori super-pixel region Neighbouring super pixels region in there is unknown super-pixel region, then by true for this non-priori super-pixel region It is set to unknown super-pixel region.

By above-mentioned steps 15 and 16, the plurality of super-pixel region is divided into priori super-pixel district Territory, non-priori super-pixel region and unknown super-pixel region, wherein, priori super-pixel region can To be expressed as foreground area, non-priori super-pixel region can be expressed as background area, unknown super picture Element region can be expressed as zone of ignorance.

Step 17, based on determined by priori super-pixel region, non-priori super-pixel region and not Know super-pixel region, generate constraint gray level image.

For example, it is possible to the pixel value of the pixel in priori super-pixel region determined by by is defined as Represent the pixel value of white, by determined by the pixel value of pixel in non-priori super-pixel region true Be set to the pixel value representing black, by determined by the pixel value of pixel in unknown super-pixel region It is defined as the pixel value representing black with white Intermediate grey, obtains retraining gray level image, this constraint Gray level image is constraint information (TriMap).

Concrete, represent that the pixel value of white is 255, represent that the pixel value of black is 0, represent black Look is 125 with the pixel value of white Intermediate grey.

Step 18, based on determined by retrain gray level image, treat stingy figure image and carry out FIG pull handle.

This step specifically can use be currently known to extract foreground target figure based on constraint gray level image The algorithm of picture, treats stingy figure image and carries out FIG pull handle.

Drawing method is scratched in the nothing supervision using the embodiment of the present invention 1 to provide, owing to generating constraint gray-scale map During Xiang, introducing super-pixel region, the position relationship between super-pixel region, angle point is convex Bag detection, and utilize the elder generation in priori super-pixel region that condition random field theoretical optimization initially determines that Test the process of label value such that it is able to determine the priori super-pixel representing foreground image more accurately Region, and then improve the accuracy automatically obtaining constraint information from figure image to be scratched.

Embodiment 2:

The embodiment of the present invention 2 provides a kind of and scratches drawing method without supervision, as in figure 2 it is shown, can include Following steps:

Step 21, treat stingy figure image and carry out super-pixel segmentation, obtain multiple super-pixel region.

In this step, the various super-pixel partitioning algorithms being currently known can be used to treat stingy figure image Carry out super-pixel segmentation, such as, use linear iteraction clustering algorithm (SLIC) to treat stingy figure image Carry out super-pixel segmentation, obtain multiple super-pixel region.

According to super-pixel segmentation result, each super-pixel region can a corresponding node, form one Individual closed loop figure.

Step 22, characteristic vector based on the plurality of super-pixel region, determine the plurality of super-pixel district Characteristic relation between each two super-pixel region in territory.

In this step, equation below specifically can be used to determine each two in the plurality of super-pixel region Characteristic relation between super-pixel region:

$E_{ij}=\left\{\begin{array}{cc}\exp (-W(i,j)\left|\right|v_i-v_j\left|{|}^2\right)& (v_j\∈C_i)\\ 0& \left(v_j\∉C_i\right)\end{array}\right.;$

Wherein, E_ijRepresent that in the plurality of super-pixel region, i-th super-pixel region and jth surpass picture Characteristic relation value between element region, v_iRepresent the characteristic vector in i-th super-pixel region, v_jTable Show the characteristic vector in jth super-pixel region, C_iRepresenting has predetermined with i-th super-pixel region The super-pixel regional ensemble of justice annexation；

W (i, j)=(| | v_i-v_k(i)|g||v_k(j)-v_j||)^-1；

Wherein, n is C_iThe quantity in the super-pixel region included, v_kI () is for meet following relational expression C_iIn the characteristic vector in a super-pixel region, this super-pixel region is in i-th super-pixel district In the neighborhood in territory, the super-pixel region nearest with other super-pixel region distances in neighborhood；

$v_k\left(i\right)=\arg \min \underset{l=1}{\overset{n}{\Σ}}\left|\right|v_k\left(i\right)-v_l\left(i\right){|}_2^2;$

Further, characteristic vector v in super-pixel region can be expressed as v=[L, a, b, R, G, B, x, y], Wherein, L, a, b represent super-pixel region each passage average under Lab color space respectively Being worth, R, G, B represent super-pixel region mean value of each passage under RGB color respectively, X, y represent the center in super-pixel region abscissa in figure image to be scratched and ordinate respectively.

Further, and the super-pixel region that i-th super-pixel region has predefined annexation, Can be the neighbouring super pixels region in i-th super-pixel region, and i-th super-pixel region In the range of two rings in the neighbouring super pixels region in neighbouring super pixels region, i.e. i-th super-pixel region Super-pixel region

Can be expressed as by the figure G after the segmentation that above-mentioned steps 21 and step 22 obtain, figure G G=(V, E), wherein, V represents the node set in figure, V_D×N={ (v₁,v₂,…,v_N-1,v_N), node with Super-pixel region one_to_one corresponding, E represents the line set between figure interior joint, is super-pixel region Between the sign of relation, the most above-mentioned E_ijThe relational matrix formed, it is a sparse matrix.

Step 23, treat stingy figure image and carry out Corner Detection, obtain the angle point collection in figure image to be scratched Close.

Strict sequencing is not had between this step 23 and above-mentioned steps 21 and step 22.

Step 24, based on the position of angle point in angle point set, determine the convex closure district that angle point set is corresponding Territory.

This step is equivalent to be attached angle point peripherally located for the position in angle point set, and even The curve connect expands outwardly, obtained convex closure region, compares straight line and connects the region model drawn a circle to approve Enclose bigger, so that initial foreground area scope determined by follow-up is the biggest, in order to reduce The final inaccuracy obtaining constraint information.

Step 25, in multiple super-pixel regions, determine the super-pixel region of convex closure region overlay to The most adjacent super-pixel region, i.e. convex closure region outward expansion one ring, the super picture of convex closure region overlay Element region and the most adjacent super-pixel region, as priori super-pixel region, wherein priori surpasses picture The priori label value in element region is 1, and the priori label value in non-priori super-pixel region is 0, and all The region that priori super-pixel region is covered is referred to as region H.

Priori label L specifically can be used to be indicated, L=[l₁,l₂,…,l_N], wherein, if i-th Individual super-pixel region belongs to region H, then l_i=1, otherwise, l_i=0.

Understand based on priori, the boundary point of the foreground target image in a sub-picture, be mostly Angle point, so, in the embodiment of the present invention 1, based on priori, by above-mentioned steps 23-25 In Corner Detection etc. process step, the region H that the priori super-pixel region determined is covered, Initial foreground area can be expressed as.

Step 26, utilize condition random field theory, priori label based on the plurality of super-pixel region Value, and characteristic relation between each two super-pixel region in the plurality of super-pixel region, to this The priori label value in multiple super-pixel regions is optimized, the priori label value after optimization be 1 super Pixel region is as optimizing priori super-pixel region.

This step, utilizes condition random field theory, the priori label value to the plurality of super-pixel region It is optimized, is i.e. equivalent to above-mentioned initial foreground area H determined is optimized.

According to condition random field theory, there is a following form:

$P\left(y\right|l)=\frac{1}{Z}\exp \{-\mathrm{\ϵ}\left(y\right|l)\};$

Being different from traditional form, take y ∈ [0,1] herein, Z is normaliztion constant, and ε (y | l) it is energy side Journey, can be written as following form:

$\mathrm{\ϵ}\left(y\right|l)=\underset{i}{\Σ}{k}_i{\left(y_i-l_i\right)}^2+\mathrm{\λ}\underset{i,j}{\Σ}\frac{1}{2}{E}_{ij}{(y_i-y_j)}^2$

Therefore, this step specifically can be in the following way:

For energy equationSeek optimal solution, obtain optimum SolveWherein, k_iRepresent the weight that i-th super-pixel region is corresponding, l_iRepresent The priori label value in i-th super-pixel region, E_ijRepresent that in multiple super-pixel region, i-th surpasses picture Characteristic relation value between element region and jth super-pixel region, λ is balance factor, for handle Hold a balance of relation between the interregional relation of super-pixel and super-pixel region and priori label value,The value of the optimal solution that expression i-th super-pixel region is corresponding, N is the number in multiple super-pixel region Amount；

Use self adaptation two value-based algorithm, to optimal solution y^*Process, obtain the plurality of super-pixel district Priori label value Y=[y after the optimization in territory₁,y₂,…,y_N], wherein, y_iRepresent i-th super-pixel district Priori label value after the optimization in territory.

Further, for energy equationAsk optimum Solve, obtain following optimal solution:

y^*=(K+ λ P)^-1(λKL)；

Wherein, P be described in the Laplacian Matrix of figure image to be scratched, K is unit matrix, and L is described The vector of the priori label value composition in multiple super-pixel regions；

P=D-E；

Wherein, D=diag{d₁₁,d₂₂,…,d_NN}；

$d_{ii}=\underset{j}{\Σ}{E}_{ij}.$

By the optimization priori super-pixel region obtained by this step, all optimization priori super-pixel districts The region that territory is covered, can be expressed as optimizing initial foreground area.

Step 27, for each optimization priori super-pixel region, if this optimization priori super-pixel district The neighbouring super pixels region in territory exists unoptimizable priori super-pixel region, then this optimization priori is surpassed Pixel region is defined as unknown super-pixel region.

This step is equivalent to, in the figure G that super-pixel segmentation obtains, be somebody's turn to do based on obtained above Priori label value Y=[y after the optimization in multiple super-pixel regions₁,y₂,…,y_N], to y_iThe region of=1 (priori super-pixel region after optimization) does the corrosion of a ring, if i.e. y_i=1, and its adjacent super picture The neighborhood that priori label value is 0, then y is there is in element in region_i=-1, priori label value is-1 expression This super-pixel region is unknown super-pixel region.

Step 28, by above-mentioned steps 27, at each optimization priori super-pixel region After reason, in this step, for each unoptimizable priori super-pixel region, if this unoptimizable is first Test and the neighbouring super pixels region in super-pixel region exists unknown super-pixel region, then by this unoptimizable Priori super-pixel region is defined as unknown super-pixel region.

This step is equivalent to, to the super-pixel region that priori label value is-1, to its priori label value Be 0 neighbouring super pixels region do the expansion of a ring, if i.e. y_i=0, and its neighbouring super pixels district Territory exists the neighborhood that priori label value is-1, then y_i=-1, priori label value be-1 represent this surpass Pixel region is unknown super-pixel region.

Thus, label y_iThen meet y_i∈-1,0,1}, corresponding, represent respectively and optimize priori super-pixel Region, unoptimizable priori super-pixel region, and unknown super-pixel region.

By above-mentioned steps 27 and step 28, the plurality of super-pixel region is divided into optimization priori Super-pixel region, unoptimizable priori super-pixel region and unknown super-pixel region, wherein, optimize Priori super-pixel region can be expressed as foreground area, and unoptimizable priori super-pixel region can represent For background area, unknown super-pixel region can be expressed as zone of ignorance.

Step 29, based on determined by priori super-pixel region, non-priori super-pixel region and not Know super-pixel region, generate constraint gray level image.

For example, it is possible to the pixel value of the pixel in priori super-pixel region determined by by is defined as Represent the pixel value of white, by determined by the pixel value of pixel in non-priori super-pixel region true Be set to the pixel value representing black, by determined by the pixel value of pixel in unknown super-pixel region It is defined as the pixel value representing black with white Intermediate grey, obtains retraining gray level image, this constraint Gray level image is constraint information (TriMap).

Concrete, represent that the pixel value of white is 255, represent that the pixel value of black is 0, represent black Look is 125 with the pixel value of white Intermediate grey.

Step 30, based on determined by retrain gray level image, treat stingy figure image and carry out FIG pull handle.

This step specifically can use be currently known to extract foreground target figure based on constraint gray level image The algorithm of picture, treats stingy figure image and carries out FIG pull handle.

Drawing method is scratched in the nothing supervision using the embodiment of the present invention 2 to provide, owing to generating constraint gray-scale map During Xiang, introducing super-pixel region, the position relationship between super-pixel region, angle point is convex Bag detection, and utilize the elder generation in priori super-pixel region that condition random field theoretical optimization initially determines that Test the process of label value such that it is able to determine the priori super-pixel representing foreground image more accurately Region, and then improve the accuracy automatically obtaining constraint information from figure image to be scratched.

Embodiment 3:

For the step 30 in the step 18 in the embodiment of the present invention 1 and the embodiment of the present invention 2, I.e. based on determined by retrain gray level image, treat stingy figure image and carry out FIG pull handle, such as Fig. 3 institute Show, the embodiment of the present invention 3 proposes following handling process, specifically may include that

Step 31, in the constraint gray level image generated, if T_fRepresent foreground area, T_bRepresent Background area, T_uRepresenting zone of ignorance, foreground area and background area may be collectively referred to as known region, Then in figure image to be scratched, based on the spy between known region pixel and zone of ignorance pixel Levy relation, be divided into meeting pre-conditioned zone of ignorance pixel in known region.

Such as, required satisfied pre-conditioned, be specifically as follows when zone of ignorance pixel i with When meeting following relation between known region pixel j, zone of ignorance pixel i is divided into In knowing the known region at area pixel point j place:

D_image(i, j) ＜ 5；

D_color(i, j) ＜ 5；

D_image(i, j) minimum for i；

Wherein, D_image(i j) represents the Euclidean distance between pixel i and pixel j, D_color(i,j) Represent the color distance between pixel i and pixel j, D_image(i j) represents satisfied for minimum In the case of above-mentioned two condition, i point is corresponding with the j point of its Euclidean distance minimum, because meeting The j point of above-mentioned two condition may have multiple.

Step 32, for each unknown pixel point, select the prospect of its correspondence, background sample pair, Specific as follows:

From unknown pixel, the N of extraction_kPaths, between path, angle isRecord is every Background dot that paths runs into first or foreground point, form prospect background pair, herein, N_k=6.

First the formula choosing the highest confidence level sample pair is:

$f(F_m,B_n)=\exp \left\{-\frac{R_d{(F_m,B_n)}^2}{{\mathrm{\σ}}^2}-\frac{\left|\right|C-F_m\left|\right|}{D_F^2}-\frac{\left|\right|C-B_m\left|\right|}{D_B^2}\right\};$

$R_d(F_{mn},B)=\frac{\left|\right|C-({\mathrm{\α}}^{\′}{F}_m+(1-{\mathrm{\α}}^{\′})B_n\left)\right||}{\left|\right|F_m-B_n\left|\right|};$

Wherein, σ is constant, (F_m,B_n) it is prospect background sample pair, D_F、D_BRepresent current pixel And the minimum range between prospect background sample pair, α ' is current pixel point, at (F_m,B_nProjection on) Estimate.

f(F_m,B_n) maximum time, corresponding (F_m,B_n) it is the highest confidence level sample pair.

Step 33, α cover as initial estimation.

After upper one group, to the estimation formulas of unknown pixel value it is: WhereinFor the prospect background sample pair that confidence level is the highest.

Definition unknown pixel value isConfidence level calculation as follows:

Wherein (F_i,B_i) it is prospect background sample pair, C_iFor Actual value in figure image to be scratched.

It is hereby achieved that the estimation to unknown pixel point transparence value:

${\widehat{\mathrm{\α}}}_i=f\left({\mathrm{\α}}_i^{*}\right){\mathrm{\α}}_i^{*}+\left(1-f\left({\mathrm{\α}}_i^{*}\right)\right)\frac{\underset{n\∈{\mathrm{\ψ}}_i}{\Σ}\[Z_{\mathrm{\α}}\left(i,j\right){\mathrm{\α}}_n^{*}\]}{\underset{n\∈{\mathrm{\ψ}}_i}{\Σ}{Z}_{\mathrm{\α}}\left(i,j\right)};$

$Z_{\mathrm{\α}}(i,j)=f\left({\mathrm{\α}}_j^{*}\right)G\left(D_{image}\right(i,j\left)\right)+\mathrm{\δ}(j\∉T_u);$

Wherein, ψ_iRepresenting the neighborhood of pixel i, G represents Gaussian function.

Each unknown pixel point transparence value is solved, finally gives α and cover the initial estimation of picture.

Step 34, α cover as double optimization.

Setting up secondary objective optimization function is:

$\mathrm{\α}={\mathrm{argmin\α}}^{T}L\mathrm{\α}+\mathrm{\λ}{(\mathrm{\α}-\widehat{\mathrm{\α}})}^{T}D(\mathrm{\α}-\widehat{\mathrm{\α}});$

Wherein, λ=100,Be expressed as initial transparent degree vector (upper step obtain), L be expressed as with Each pixel is Laplce's battle array of the figure correspondence of node composition.Concrete calculation is as follows:

Wherein, W_ijFor i point in image and the distance weights of j point, calculation is as follows:

$W_{ij}=\underset{k}{\overset{\left(i,j\right)\∈Wk}{\Σ}}\frac{1}{9}\left(1+\left(C_i-{\mathrm{\μ}}_k\right){\left({\Σ}_k\frac{\mathrm{\ϵ}}{9}I\right)}^{-1}\left(C_j-{\mathrm{\μ}}_k\right)\right);$

Wherein, Wk represents the window pixel collection comprising i and j, is 3*3 herein；μ_kFor window Average；Σ_kRepresent that window standard is poor；ε=10^-5。

By solving quadratic objective function, i.e. can get α and cover picture, and then complete foreground target image Extract.

Embodiment 4:

Based on same inventive concept, scratch figure side according to what the above embodiment of the present invention 1 provided without supervision Method, correspondingly, the embodiment of the present invention 4 additionally provides a kind of without the stingy map device of supervision, and its structure is shown It is intended to as shown in Figure 4, specifically include:

First segmentation module 41, is used for treating stingy figure image and carries out super-pixel segmentation, obtains multiple super Pixel region；

First Corner Detection module 42, for described figure image to be scratched is carried out Corner Detection, obtains Angle point set in described figure image to be scratched；

First convex closure area determination module 43, is used for based on the position of angle point in described angle point set, Determine the convex closure region that described angle point set is corresponding；

First expansion module 44, in the plurality of super-pixel region, determines described convex closure district The most adjacent super-pixel region, super-pixel region that territory covers, the super picture of described convex closure region overlay Element region and the most adjacent described super-pixel region, as priori super-pixel region；

First area determines module 45, is used for for each priori super-pixel region, if this priori The neighbouring super pixels region in super-pixel region exists non-priori super-pixel region, then this priori is surpassed Pixel region is defined as unknown super-pixel region；

Second area determines module 46, for for each non-priori super-pixel region, if this is non- The neighbouring super pixels region in priori super-pixel region exists unknown super-pixel region, then by this non-elder generation Test super-pixel region and be defined as unknown super-pixel region；

First constraint information generation module 47, for based on determined by priori super-pixel region, non- Priori super-pixel region and unknown super-pixel region, generate constraint gray level image；

First FIG pull handle module 48, for based on described constraint gray level image, scratching figure to described waiting Image carries out FIG pull handle.

The function of above-mentioned each unit may correspond to the respective handling step in flow process shown in Fig. 1, at this Repeat no more.

Embodiment 5:

Based on same inventive concept, scratch figure side according to what the above embodiment of the present invention 2 provided without supervision Method, correspondingly, the embodiment of the present invention 5 additionally provides a kind of without the stingy map device of supervision, and its structure is shown It is intended to as it is shown in figure 5, specifically include:

Second segmentation module 51, is used for treating stingy figure image and carries out super-pixel segmentation, obtains multiple super Pixel region；

Characteristic relation determines module 52, for characteristic vector based on the plurality of super-pixel region, Determine the characteristic relation between each two super-pixel region in the plurality of super-pixel region；

Second Corner Detection module 53, for described figure image to be scratched is carried out Corner Detection, obtains Angle point set in described figure image to be scratched；

Second convex closure area determination module 54, is used for based on the position of angle point in described angle point set, Determine the convex closure region that described angle point set is corresponding；

Second expansion module 55, in the plurality of super-pixel region, determines described convex closure district The most adjacent super-pixel region, super-pixel region that territory covers, the super picture of described convex closure region overlay Element region and the most adjacent described super-pixel region, as priori super-pixel region, wherein priori The priori label value in super-pixel region is 1, and the priori label value in non-priori super-pixel region is 0；

Optimize module 56, be used for utilizing condition random field theory, based on the plurality of super-pixel region Priori label value, and spy between each two super-pixel region in the plurality of super-pixel region Levy relation, the priori label value in the plurality of super-pixel region is optimized, the priori after optimization Label value is that the super-pixel region of 1 is as optimizing priori super-pixel region；

3rd area determination module 57, for for each optimization priori super-pixel region, if should Optimize in the neighbouring super pixels region in priori super-pixel region and there is unoptimizable priori super-pixel region, Then this optimization priori super-pixel region is defined as unknown super-pixel region；

4th area determination module 58, is used for for each unoptimizable priori super-pixel region, if The neighbouring super pixels region in this unoptimizable priori super-pixel region exists unknown super-pixel region, then This unoptimizable priori super-pixel region is defined as unknown super-pixel region；

Second constraint information generation module 59, for based on determined by optimize priori super-pixel region, Unoptimizable priori super-pixel region and unknown super-pixel region, generate constraint gray level image；

Second FIG pull handle module 60, for based on described constraint gray level image, scratching figure to described waiting Image carries out FIG pull handle.

The function of above-mentioned each unit may correspond to the respective handling step in flow process shown in Fig. 2, at this Repeat no more.

What embodiments herein was provided can be realized by computer program without the stingy map device of supervision. Those skilled in the art are it should be appreciated that above-mentioned Module Division mode is only numerous Module Division One in mode, if being divided into other modules or not dividing module, as long as scratching figure dress without supervision Put and there is above-mentioned functions, all should be within the protection domain of the application.

The application is with reference to method, equipment (system) and the computer journey according to the embodiment of the present application The flow chart of sequence product and/or block diagram describe.It should be understood that can be real by computer program instructions Show each flow process in flow chart and/or block diagram and/or square frame and flow chart and/or side Flow process in block diagram and/or the combination of square frame.These computer program instructions can be provided to general meter The processor of calculation machine, special-purpose computer, Embedded Processor or other programmable data processing device To produce a machine so that by computer or the processor of other programmable data processing device The instruction performed produces for realizing at one flow process of flow chart or multiple flow process and/or block diagram one The device of the function specified in individual square frame or multiple square frame.

These computer program instructions may be alternatively stored in and can guide at computer or other programmable datas In the computer-readable memory that reason equipment works in a specific way so that being stored in this computer can The instruction read in memory produces the manufacture including command device, and this command device realizes in flow process The merit specified in one flow process of figure or multiple flow process and/or one square frame of block diagram or multiple square frame Energy.

These computer program instructions also can be loaded into computer or other programmable data processing device On so that on computer or other programmable devices, perform sequence of operations step to produce calculating The process that machine realizes, thus the instruction performed on computer or other programmable devices provides and is used for Realize in one flow process of flow chart or multiple flow process and/or one square frame of block diagram or multiple square frame The step of the function specified.

Obviously, those skilled in the art can carry out various change and modification to the present invention and not take off From the spirit and scope of the present invention.So, if these amendments of the present invention and modification belong to this Within the scope of bright claim and equivalent technologies thereof, then the present invention be also intended to comprise these change and Including modification.

Claims (10)

1. scratch drawing method without supervision for one kind, it is characterised in that including:

Treat stingy figure image and carry out super-pixel segmentation, obtain multiple super-pixel region；

Described figure image to be scratched is carried out Corner Detection, obtain described in angle point collection in figure image to be scratched Close；

Based on the position of angle point in described angle point set, determine the convex closure district that described angle point set is corresponding Territory；

In the plurality of super-pixel region, determine that the super-pixel region of described convex closure region overlay is outside Adjacent super-pixel region, the super-pixel region of described convex closure region overlay is the most adjacent with described Super-pixel region, as priori super-pixel region；

For each priori super-pixel region, if the neighbouring super pixels region in this priori super-pixel region In there is non-priori super-pixel region, then this priori super-pixel region is defined as unknown super-pixel Region；

For each non-priori super-pixel region, if the neighbouring super pixels in this non-priori super-pixel region Region exists unknown super-pixel region, then this non-priori super-pixel region is defined as unknown super Pixel region；

Priori super-pixel region, non-priori super-pixel region and unknown super-pixel district determined by based on Territory, generates constraint gray level image；

Based on described constraint gray level image, described figure image to be scratched is carried out FIG pull handle.

2. the method for claim 1, it is characterised in that priori super-pixel district determined by based on Territory, non-priori super-pixel region and unknown super-pixel region, generate constraint gray level image, bag Include:

The pixel value of the pixel in priori super-pixel region determined by by is defined as representing white Pixel value, by determined by the pixel value of pixel in non-priori super-pixel region be defined as representing The pixel value of black, by determined by the pixel value of pixel in unknown super-pixel region be defined as Represent the pixel value of black and white Intermediate grey, obtain retraining gray level image.

3. scratch drawing method without supervision for one kind, it is characterised in that including:

Treat stingy figure image and carry out super-pixel segmentation, obtain multiple super-pixel region；

Characteristic vector based on the plurality of super-pixel region, determines in the plurality of super-pixel region every Characteristic relation between two super-pixel regions；

Described figure image to be scratched is carried out Corner Detection, obtain described in angle point collection in figure image to be scratched Close；

Based on the position of angle point in described angle point set, determine the convex closure district that described angle point set is corresponding Territory；

In the plurality of super-pixel region, determine that the super-pixel region of described convex closure region overlay is outside Adjacent super-pixel region, the super-pixel region of described convex closure region overlay is the most adjacent with described Super-pixel region, as priori super-pixel region, the wherein first standard inspection in priori super-pixel region Label value is 1, and the priori label value in non-priori super-pixel region is 0；

Utilize condition random field theory, priori label value based on the plurality of super-pixel region, and Characteristic relation between each two super-pixel region in the plurality of super-pixel region, to described many The priori label value in individual super-pixel region is optimized, the priori label value after optimization be 1 super Pixel region is as optimizing priori super-pixel region；

For each optimization priori super-pixel region, if this optimization priori super-pixel region is adjacent super Pixel region exists unoptimizable priori super-pixel region, then by this optimization priori super-pixel region It is defined as unknown super-pixel region；

For each unoptimizable priori super-pixel region, if the phase in this unoptimizable priori super-pixel region Adjacent super-pixel region exists unknown super-pixel region, then by this unoptimizable priori super-pixel region It is defined as unknown super-pixel region；

Priori super-pixel region, unoptimizable priori super-pixel region and the unknown is optimized determined by based on Super-pixel region, generates constraint gray level image；

Based on described constraint gray level image, described figure image to be scratched is carried out FIG pull handle.

4. method as claimed in claim 3, it is characterised in that based on the plurality of super-pixel region Characteristic vector, determines the feature between each two super-pixel region in the plurality of super-pixel region Relation, including:

Equation below is used to determine in the plurality of super-pixel region between each two super-pixel region Characteristic relation:

$E_{ij}=\left\{\begin{array}{cc}\exp (-W(i,j\left)\right||v_i-v_j|{|}^2)& (v_j\∉C_i)\\ 0& (v_j\∈C_i)\end{array}\right.;$

Wherein, E_ijRepresent that in the plurality of super-pixel region, i-th super-pixel region and jth surpass picture Characteristic relation value between element region, v_iRepresent the characteristic vector in i-th super-pixel region, v_jTable Show the characteristic vector in jth super-pixel region, C_iRepresenting has pre-with i-th super-pixel region The super-pixel regional ensemble of definition annexation；

W (i, j)=(| | v_i-v_k(i)||g||v_k(j)-v_j||)^-1；

Wherein, n is C_iThe quantity in the super-pixel region included, v_kI () is the C meeting following relational expression_i In the characteristic vector in a super-pixel region；

$v_k\left(i\right)=\arg \min \underset{l=1}{\overset{n}{\Σ}}\left|\right|v_k\left(i\right)-v_l\left(i\right)|{|}_2^2.$

5. method as claimed in claim 4, it is characterised in that the characteristic vector v table in super-pixel region Being shown as v=[L, a, b, R, G, B, x, y], wherein, L, a, b represent that super-pixel region is at Lab respectively The mean value of each passage under color space, R, G, B represent that super-pixel region is at RGB respectively The mean value of each passage under color space, x, y represent the centre bit in super-pixel region respectively Put the abscissa in described figure image to be scratched and ordinate.

6. method as claimed in claim 3, it is characterised in that utilize condition random field theory, based on In the priori label value in the plurality of super-pixel region, and the plurality of super-pixel region every two Characteristic relation between individual super-pixel region, the priori label value to the plurality of super-pixel region It is optimized, including:

For energy equationSeek optimal solution, obtain optimum SolveWherein, k_iRepresent the weight that i-th super-pixel region is corresponding, l_iTable Show the priori label value in i-th super-pixel region, E_ijRepresent in the plurality of super-pixel region Characteristic relation value between i super-pixel region and jth super-pixel region, λ for balance because of Son,The value of the optimal solution that expression i-th super-pixel region is corresponding, N is the plurality of super picture The quantity in element region；

Use self adaptation two value-based algorithm, to described optimal solution y^*Process, obtain the plurality of super picture Priori label value Y=[y after the optimization in element region₁,y₂,…,y_N], wherein, y_iRepresent that i-th surpasses Priori label value after the optimization of pixel region.

7. method as claimed in claim 6, it is characterised in that for energy equationSeek optimal solution, obtain following optimal solution:

y^*=(K+ λ P)^-1(λKL)；

Wherein, P be described in the Laplacian Matrix of figure image to be scratched, K is unit matrix, and L is described The vector of the priori label value composition in multiple super-pixel regions；

P=D-E；

Wherein, D=diag{d₁₁,d₂₂,…,d_NN}；

$d_{ii}=\underset{j}{\Σ}{E}_{ij}.$

8. method as claimed in claim 3, it is characterised in that optimize priori determined by based on and surpass picture Element region, unoptimizable priori super-pixel region and unknown super-pixel region, generate constraint gray scale Image, including:

The pixel value optimizing the pixel in priori super-pixel region determined by by is defined as representing white The pixel value of look, by determined by the pixel value of pixel in unoptimizable priori super-pixel region true Be set to the pixel value representing black, by determined by the pixel of pixel in unknown super-pixel region Value is defined as the pixel value representing black with white Intermediate grey, obtains retraining gray level image.

9. scratch map device without supervision for one kind, it is characterised in that including:

First segmentation module, is used for treating stingy figure image and carries out super-pixel segmentation, obtain multiple super-pixel Region；

First Corner Detection module, for described figure image to be scratched is carried out Corner Detection, obtains described Angle point set in figure image to be scratched；

First convex closure area determination module, for based on the position of angle point in described angle point set, determines The convex closure region that described angle point set is corresponding；

First expansion module, in the plurality of super-pixel region, determines that described convex closure region is covered The most adjacent super-pixel region, super-pixel region of lid, the super-pixel of described convex closure region overlay Region and the most adjacent described super-pixel region, as priori super-pixel region；

First area determines module, for for each priori super-pixel region, if this priori surpasses picture There is non-priori super-pixel region in the neighbouring super pixels region in element region, then this priori is surpassed picture Element region is defined as unknown super-pixel region；

Second area determines module, is used for for each non-priori super-pixel region, if this non-priori The neighbouring super pixels region in super-pixel region exists unknown super-pixel region, then by this non-priori Super-pixel region is defined as unknown super-pixel region；

First constraint information generation module, for based on determined by priori super-pixel region, non-priori Super-pixel region and unknown super-pixel region, generate constraint gray level image；

First FIG pull handle module, for based on described constraint gray level image, to described figure image to be scratched Carry out FIG pull handle.

10. scratch map device without supervision for one kind, it is characterised in that including:

Second segmentation module, is used for treating stingy figure image and carries out super-pixel segmentation, obtain multiple super-pixel Region；

Characteristic relation determines module, for characteristic vector based on the plurality of super-pixel region, determines Characteristic relation between each two super-pixel region in the plurality of super-pixel region；

Second Corner Detection module, for described figure image to be scratched is carried out Corner Detection, obtains described Angle point set in figure image to be scratched；

Second convex closure area determination module, for based on the position of angle point in described angle point set, determines The convex closure region that described angle point set is corresponding；

Second expansion module, in the plurality of super-pixel region, determines that described convex closure region is covered The most adjacent super-pixel region, super-pixel region of lid, the super-pixel of described convex closure region overlay Region and the most adjacent described super-pixel region, as priori super-pixel region, wherein priori The priori label value in super-pixel region is 1, and the priori label value in non-priori super-pixel region is 0；

Optimize module, be used for utilizing condition random field theory, elder generation based on the plurality of super-pixel region Test the feature between each two super-pixel region in label value, and the plurality of super-pixel region Relation, is optimized the priori label value in the plurality of super-pixel region, the priori after optimization Label value is that the super-pixel region of 1 is as optimizing priori super-pixel region；

3rd area determination module, is used for for each optimization priori super-pixel region, if this optimization The neighbouring super pixels region in priori super-pixel region exists unoptimizable priori super-pixel region, then This optimization priori super-pixel region is defined as unknown super-pixel region；

4th area determination module, for for each unoptimizable priori super-pixel region, if this is non- Optimize and the neighbouring super pixels region in priori super-pixel region exists unknown super-pixel region, then will This unoptimizable priori super-pixel region is defined as unknown super-pixel region；

Second constraint information generation module, for based on determined by optimize priori super-pixel region, non- Optimize priori super-pixel region and unknown super-pixel region, generate constraint gray level image；

Second FIG pull handle module, for based on described constraint gray level image, to described figure image to be scratched Carry out FIG pull handle.