CN105809672B - A kind of image multiple target collaboration dividing method constrained based on super-pixel and structuring - Google Patents
A kind of image multiple target collaboration dividing method constrained based on super-pixel and structuring Download PDFInfo
- Publication number
- CN105809672B CN105809672B CN201610120407.7A CN201610120407A CN105809672B CN 105809672 B CN105809672 B CN 105809672B CN 201610120407 A CN201610120407 A CN 201610120407A CN 105809672 B CN105809672 B CN 105809672B
- Authority
- CN
- China
- Prior art keywords
- super
- pixel
- segmentation
- subtree
- target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 51
- 230000011218 segmentation Effects 0.000 claims abstract description 105
- 238000005457 optimization Methods 0.000 claims abstract description 12
- 239000013598 vector Substances 0.000 claims description 16
- 238000011156 evaluation Methods 0.000 claims description 15
- 238000012545 processing Methods 0.000 claims description 8
- 238000005070 sampling Methods 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 5
- 238000013139 quantization Methods 0.000 claims description 5
- 238000009795 derivation Methods 0.000 claims description 4
- 238000007476 Maximum Likelihood Methods 0.000 claims description 3
- 238000012217 deletion Methods 0.000 claims description 3
- 230000037430 deletion Effects 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 238000012706 support-vector machine Methods 0.000 claims description 3
- 238000012549 training Methods 0.000 claims description 3
- 238000001514 detection method Methods 0.000 abstract description 4
- 230000007246 mechanism Effects 0.000 abstract description 3
- 238000013480 data collection Methods 0.000 abstract 1
- 238000005259 measurement Methods 0.000 description 6
- 238000011160 research Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 235000013399 edible fruits Nutrition 0.000 description 4
- 230000001154 acute effect Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000012935 Averaging Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000005286 illumination Methods 0.000 description 2
- 238000003709 image segmentation Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Abstract
Non-supervisory dividing method is cooperateed with the image multiple target that structuring constrains based on super-pixel the invention discloses a kind of.It include several image data collection of common objects for one group, and every width picture may include multiple common objects, this method can accurately be partitioned into common objects.Firstly, this method carries out pre-segmentation operation to the image set of input, the image after obtaining over-segmentation;Then, the classification of foreground and background is carried out to all super-pixel based on target detection mechanism, background class device before study obtains;Finally, based on classifier obtain as a result, modeled to foreground target, assume to complete the Accurate Segmentation to target using the algorithm of Combinatorial Optimization with tree graph constraint using forest model.Compared with analogous algorithms, the present invention is assumed by proposing new forest model and method for solving, passes through the constrained optimization combinatorial optimization algorithm of tree graph and improves segmentation precision, can adapt to various complex scenes.
Description
Technical field
The present invention relates to a kind of image multiple targets constrained based on super-pixel and structuring to cooperate with dividing method, is suitable for figure
The fields such as multiple target collaboration segmentation, the object segmentation in sports picture and the picture classification identification of piece.
Background technique
In computer vision field, image segmentation is a basic and classical problem, its solution can be to other
Numerous image processing problems such as target identification, object classification play good booster action.In practical applications, intelligence prison
The fields such as control, medical diagnosis, robot technology and intelligence machine, industrial automation or even military guidance all have with image segmentation
Closely connection.By means of internet, it includes same object or same category object that people, which can be very easy to obtain,
A large amount of pictures, and how to be distinguished automatically from this kind of picture and be partitioned into what the interested common objects of people were studied as us
Main purpose.It can be partitioned into interested target by the bottom-up information (color, texture etc.) of image, but only rely only on bottom
The image data information of layer can not obtain desired segmentation result, and the implicit information across picture can then help to distinguish what is
The common objects for needing to recognize.It is this kind of completed using the plurality of pictures comprising same object or the same category object to feel it is emerging
The research that the common objects of interest are split, referred to as collaboration segmentation.Collaboration segmentation is the popular research risen in recent years
Theme, the more research work about collaboration segmentation existing at present.However, make a general survey of in relation to cooperate with segmentation field research and
Using it is found that current collaboration segmentation area research remains unchanged, there are many technical problems are as follows:
1) existing method mainly utilizes the features such as the color of bottom, shape, and has ignored the high-rise base that can be learnt
The structuring constraint of object under the feature and multiple target scene of super-pixel;
2) current mainstream algorithm is directed to single goal segmentation design mostly, and to the segmentation of multiple target, often effect is undesirable,
It has no and targetedly optimizes;
3) scalability of most methods is undesirable, can not solve the processing to large database concept.
Above technical problem is that cutting techniques is cooperateed with to bring many puzzlements in the extensive use of MultiMedia Field, is developed
A set of method application value with higher suitable for multiple target collaboration segmentation out.
Summary of the invention
In order to solve problem existing in the prior art, constrained the invention discloses a kind of based on super-pixel and structuring
Image multiple target cooperates with dividing method, and this method is suitable for the segmentation of the common objects with multiple target, passes through target detection machine
The background class device before the method that study combines obtains is made, so that algorithm has better scalability.And the forest proposed
Model and the iterative splitting algorithm based on tree graph structuring constraint, effectively solve Combinatorial Optimization energy model, thus
So that the segmentation to multiple target is more accurate, and substantially increase computational efficiency.
The invention adopts the following technical scheme: a kind of image multiple target constrained based on super-pixel and structuring cooperates with segmentation
Method comprises the steps of:
(1) image pre-segmentation: for the image data set I={ I comprising common objective object1..., INIn each width
Image Ii, i=1,2 ..., N carry out over-segmentation processing, obtain super-pixel collection
(2) automatic target is found: the super-pixel collection based on each imageCount each super-pixelConspicuousness
ValueWith repeated value wim, and calculate super-pixelEvaluation of estimate scoreim, It will
Evaluation of estimate is less than 0.6 × max (scorei) super-pixel be set as background, by evaluation of estimate be more than or equal to 0.6 × max (scorei)
Super-pixel be set as prospect;max(scorei) it is super-pixel collectionThe evaluation of estimate of the middle maximum super-pixel of evaluation of estimate;
(3) classifier learns.It is found by automatic target by the super-pixel collection in training setIt is divided into foreground and background,
For each super-pixel, described using the characteristic vector of 2004 following dimensions: (a) hsv color of 800 Dimension Vector Quantization of Linear Prediction is indicated
(k mean cluster obtains);(b) SIFT bag of words (1200 dimensions, respectively with 16,24,32 pixels that multiple dimensioned intensive sampling obtains
It is sampled for the image block on side, the sampling interval is 3 pixels);(c) 4 binaryzation features, to describe super-pixel and four sides of image
The contact situation on boundary.Based on features above, background class device before being obtained using the support vector machines learning method of standard.
(4) Target Modeling: being based on step (2) sorted information, is established to common objective object based on hsv color space
Object module ΨfWith background model Ψb.Using Hellinger distance metric method calculate separately super-pixel or super-pixel combination with
Similarity degree between object moduleSimilarity degree between super-pixel or super-pixel combination and background model
Object module ΨfMethod for building up it is as follows: by original image carry out color space transformation, obtain hsv color space
Under image;To under hsv color space image H, S, V and " G " four color components carry out uniform quantizations, count target
Distribution of the object on each color component obtains histogram distribution, i.e. object module Ψf;By the same way, background is counted
Distribution of the image on each color component obtains histogram distribution, i.e. background model Ψb;Wherein " G " component represents saturation degree
The color quantizing value of pixel lower than 5%;
Be respectively as follows:The normalization for being
Value, the normalized value for being, and
C is all section numbers after equal part, hRAnd hfThe face of super-pixel or super-pixel combination R after respectively normalizing
The color histogram of Color Histogram and object module, hR′And hbThe face of super-pixel or super-pixel combination R ' after respectively normalizing
The color histogram of Color Histogram and background model.
(5) based on the segmentation of super-pixel: utilizing object module ΨfWith background model Ψb, using the algorithm pair of Combinatorial Optimization
The subseries again of background before super-pixel carries out, to obtain the final segmentation of target object;It is proposed forest model it is assumed that i.e. vacation
If each super-pixel correspond to a vertex, for single goal divide, last segmentation result by multiple adjoinings super-pixel structure
At, and adjacent map can be expressed asSubtree;For Segmentation of Multi-target, last segmentation result is represented by adjacent map's
The forest that multiple subtrees are constituted.To sum up, it is assumed that last segmentation result is adjacent mapMultiple subtrees constitute forest.Pass through
Establish adjacent mapTo infer that the method for subtree collection determines last segmentation result;Specific implementation process
It is as follows:
(5.1) adjacent map is constructed: assuming that each super-pixel in image corresponds to a vertex in figure, it is two adjacent
It is connected between super-pixel by a line, thus constitutes adjacent mapKnot is divided for final target object
Fruit, it is assumed that the forest that its multiple subtree for including by adjacent map is constituted;
(5.2) it establishes numerical model solution: establishing numerical model, the problem of Target Segmentation is converted into combinatorial optimization problem
Solution, it is as follows:
When R is super-pixel or super-pixel combination in prospect,When R ' is super-pixel or super picture in background
When element combination,Constraint condition indicates one kind before can only belonging to for any one super-pixel R in background.Pass through
Derivation can obtain, and to solution segmentation result, actually can be exchanged into the method for solving optimal subtree collection, and require optimal subtree
Set, needs first to estimate maximum spanning tree;
(5.3) it derives maximum spanning tree: obtaining all possible candidate by the beam search method of beam search
Subtree collectionBased on candidate subtree collectionMaximum spanning tree is obtained by the method for maximal possibility estimationIt derives such as
Under:
Indicate all potential spanning tree set,It indicates data likelihood probability, can finally export,
Candidate subtree collection,For a certain subtree,Expression pairMaximum likelihood
Estimation, δ () are indicator function, δ ((x, y) ∈ Cq) indicate whether side (x, y) belongs to a certain subtree Cq;
For subtree CqWith the similarity degree of object module, indicating whether side belongs to a certain subtree, P (x, y) indicates the probability of side (x, y),For the maximal possibility estimation to P (x, y).Maximum spanning tree can be obtained by above formulaMaximal possibility estimation.
(5.4) search segmentation subtree collection: it is based on maximum spanning treeMaximal possibility estimation acquireThen pass through
Dynamic programming techniques existMiddle search obtains optimal subtree collection, and the specific implementation steps are as follows:
(5.4.1) is for image Ii, preceding background class device is acted on into super-pixel setObtain the kind for being classified as prospect
Sub- super-pixel setIts seed super-pixel setSurpassed by discrete seed
Pixel is constituted, and is ranked up to obtain according to the similarity degree of each seed super-pixel and object module first
(5.4.2) chooses the super-pixel s closest to object module1As start node, infer maximum spanning tree simultaneously with this
Obtain corresponding optimal subtree and its corresponding segmentation resultJudge the similarity degree of this segmentation result and object module: such as
Fruit similarity degree is eligibleThen think that segmentation result is effective, otherwise willIt is set as empty setAnd the wrong seed super-pixel for including in segmentation area is fed back toCarry out deletion update;
(5.4.3) traversal setFind out the segmentation result corresponding to optimal subtree before
It whether there is seed super-pixel s other than regionk, then repeat if it exists more than step obtain segmentation resultSimilarly carry out with
The similarity of object module judges and subsequent processing, updates segmentation resultWith seed super-pixel set;
(5.4.4) is completed to seed super-pixel setWhole traversals after, we obtain finally being directed to image Ii's
Segmentation resultWith updated seed super-pixel setAnd object module is completed according to these information
Update and seed super-pixel constraint information update, to make the estimation of model more close to change present in real scene
The seed super-pixel for changing situation and debug, then begins to iteration next time.
(6) iterative segmentation: the object module in step 4 is updated according to the segmentation result that step 5 obtains, according to step 5 institute
The method stated, then be split;
(7) step 6 is repeated, until final segmentation result no longer changes to arrive final segmentation result.
Further, in the step 2, super-pixel significance valueMeasurement specifically:
By conspicuousness detection technique, to the i-th width image IiObtain original Saliency maps φi, then calculate each super-pixel
For the conspicuousness mean value for all pixels point for including as its measurement, specific calculating is as follows:
WhereinI-th width image IiIn m-th of super-pixel RimAverage significance value,Indicate j-th of pixel
Significance value, area (Rim) it is super-pixel RimThe number of pixels for including.
Super-pixel repeatability value wimMeasurement, specifically:
Each super-pixel minimum value at a distance from all super-pixel in other each images is measured, N-1 most narrow spacings are obtained
From { d (Rim, Ik)}k≠i, further to N-1 minimum range { d (Rim, Ik)}k≠iIt is averaging, obtains average minimumIts
Middle distance metric d (Rim, Ik) distance weighted by the vector distance based on hsv color and the bag of words based on SIFT feature
It arrives, specific as follows:
Wherein cimAnd gimRespectively represent the i-th width image IiIn m-th of super-pixel RimHsv color characteristic vector and SIFT
Bag of words characteristic vector, ckm′And gkm′Respectively represent kth width image IkIn a super-pixel R of m 'km′Hsv color Characteristic Vectors
Amount and SIFT bag of words characteristic vector;
Super-pixel repeatability metric weights w is calculated by sigmoid formulaim:
Wherein μ and σ is the parameter for controlling the sigmoid functional form, μ=0.5, σ=0.1.
Further, step 6 specifically:
(6.1) according to newest segmentation result, foreground target model before updating is allowed to be more nearly mesh to be split
Mark;
(6.2) it according to updated object module, regenerates all possible candidate subtree collection and estimates maximum
Spanning tree;
(6.3) it according to updated object module and maximum spanning tree, is searched for again using dynamic programming techniques by subtree
Gather the forest constituted, obtains segmentation result;
(6.4) judge whether to meet cut-off condition, i.e., whether last segmentation result no longer changes.If satisfied, then iteration knot
Beam;If not satisfied, then repeating (6.1)-(6.3).
The invention adopts the above technical scheme compared with prior art, has following technical effect that
1) combining target auto discovery mechanism and study, the present invention can obtain a general preceding background class device, right
The scalability of big data image set is preferable.
2) the seed super-pixel set obtained based on preceding background class device is more directly obtained by target auto discovery mechanism
Prospect it is more accurate, help to improve subsequent segmentation precision.
3) forest model and tree graph structuring constraint condition are proposed, segmentation accuracy greatly improved, especially for having
The Segmentation of Multi-target effect of complicated fine structure is preferable, and asks optimal solution to provide new optimization for the equation of Combinatorial Optimization
Derivation algorithm.
Detailed description of the invention
Fig. 1 is overall flow figure of the invention;
Fig. 2 is preceding background class device learning process schematic diagram;
Fig. 3 is the schematic diagram based on super-pixel segmentation;
Fig. 4 is the Segmentation of Multi-target result in the case of scale, posture acute variation;
Specific embodiment
With reference to the accompanying drawing by specific embodiment, technical solution of the present invention is described in further detail.
Following embodiment is implemented under the premise of the technical scheme of the present invention, gives detailed embodiment and tool
The operating process of body, but protection scope of the present invention is not limited to following embodiments.
The present embodiment handles the multiclass image in disclosed iCoseg data set.The image of these classifications there is
The acute variations such as color, illumination condition, posture, scale, and the case where there are multiple common objects in image, give existing segmentation
Technology brings huge challenge.Fig. 1 is overall flow figure of the invention, and Fig. 2 is classifier learning process schematic diagram, and Fig. 3 is base
In the schematic diagram of super-pixel segmentation.The present embodiment comprises the steps of:
(1) image pre-segmentation: for the image data set I={ I comprising common objective object1..., INIn each width
Image Ii, i=1,2 ..., N carry out over-segmentation processing, obtain super-pixel collection
(2) automatic target is found: the super-pixel collection based on each imageCount each super-pixelConspicuousness
ValueWith repeated value wim, and calculate super-pixelEvaluation of estimate scoreim, It will
Evaluation of estimate is less than 0.6 × max (scorei) super-pixel be set as background, by evaluation of estimate be more than or equal to 0.6 × max (scorei)
Super-pixel be set as prospect;max(scorei) it is super-pixel collectionThe evaluation of estimate of the middle maximum super-pixel of evaluation of estimate;
(3) classifier learns.It is found by automatic target by the super-pixel collection in training setIt is divided into foreground and background, it is right
In each super-pixel, described using the characteristic vector of 2004 following dimensions: (a) hsv color of 800 Dimension Vector Quantization of Linear Prediction indicates (k
Mean cluster obtains);(b) (1200 tie up the SIFT bag of words that multiple dimensioned intensive sampling obtains, and are with 16,24,32 pixels respectively
The image block on side samples, and the sampling interval is 3 pixels);(c) 4 binaryzation features, to describe super-pixel and four boundaries of image
Contact situation.Based on features above, background class device before being obtained using the support vector machines learning method of standard.
(4) Target Modeling: being based on step (2) sorted information, is established to common objective object based on hsv color space
Object module ΨfWith background model Ψb.Using Hellinger distance metric method calculate separately super-pixel or super-pixel combination with
Similarity degree between object moduleSimilarity degree between super-pixel or super-pixel combination and background model
Object module ΨfMethod for building up it is as follows: by original image carry out color space transformation, obtain hsv color space
Under image;To under hsv color space image H, S, V and " G " four color components carry out uniform quantizations, count target
Distribution of the object on each color component obtains histogram distribution, i.e. object module Ψf;By the same way, background is counted
Distribution of the image on each color component obtains histogram distribution, i.e. background model Ψb;Wherein " G " component represents saturation degree
The color quantizing value of pixel lower than 5%;
Be respectively as follows:The normalization for being
Value, the normalized value for being, and
C is all section numbers after equal part, hRAnd hfThe face of super-pixel or super-pixel combination R after respectively normalizing
The color histogram of Color Histogram and object module, hR′And hbThe face of super-pixel or super-pixel combination R ' after respectively normalizing
The color histogram of Color Histogram and background model.
(5) based on the segmentation of super-pixel: utilizing object module ΨfWith background model Ψb, using the algorithm pair of Combinatorial Optimization
The subseries again of background before super-pixel carries out, to obtain the final segmentation of target object;It is proposed forest model it is assumed that i.e. vacation
If each super-pixel correspond to a vertex, for single goal divide, last segmentation result by multiple adjoinings super-pixel structure
At, and adjacent map can be expressed asSubtree;For Segmentation of Multi-target, last segmentation result is represented by adjacent map
Multiple subtrees constitute forest.To sum up, it is assumed that last segmentation result is adjacent mapMultiple subtrees constitute forest.It is logical
It crosses and establishes adjacent mapTo infer that the method for subtree collection determines last segmentation result;It implemented
Journey is as follows:
(5.1) adjacent map is constructed: assuming that each super-pixel in image corresponds to a vertex in figure, it is two adjacent
It is connected between super-pixel by a line, thus constitutes adjacent mapKnot is divided for final target object
Fruit, it is assumed that the forest that its multiple subtree for including by adjacent map is constituted;
(5.2) it establishes numerical model solution: establishing numerical model, the problem of Target Segmentation is converted into combinatorial optimization problem
Solution, it is as follows:
When R is super-pixel or super-pixel combination in prospect,When R ' is super-pixel or super picture in background
When element combination,Constraint condition indicates one kind before can only belonging to for any one super-pixel R in background.Pass through
Derivation can obtain, and to solution segmentation result, actually can be exchanged into the method for solving optimal subtree collection, and require optimal subtree
Set, needs first to estimate maximum spanning tree;
(5.3) it derives maximum spanning tree: obtaining all possible candidate by the beam search method of beam search
Subtree collectionBased on candidate subtree collectionMaximum spanning tree is obtained by the method for maximal possibility estimationIt derives such as
Under:
Indicate all potential spanning tree set,It indicates data likelihood probability, can finally export,
Candidate subtree collection,For a certain subtree,Expression pairMaximum likelihood
Estimation, δ () are indicator function, δ ((x, y) ∈ Cq) indicate whether side (x, y) belongs to a certain subtree Cq;
For subtree CqWith the similarity degree of object module, indicating whether side belongs to a certain subtree, P (x, y) indicates the probability of side (x, y),For the maximal possibility estimation to P (x, y).Maximum spanning tree can be obtained by above formulaMaximal possibility estimation.
(5.4) search segmentation subtree collection: it is based on maximum spanning treeMaximal possibility estimation acquireThen pass through
Dynamic programming techniques existMiddle search obtains optimal subtree collection, and the specific implementation steps are as follows:
(5.4.1) is for image Ii, preceding background class device is acted on into super-pixel setObtain the kind for being classified as prospect
Sub- super-pixel setIts seed super-pixel setSurpassed by discrete seed
Pixel is constituted, and is ranked up to obtain according to the similarity degree of each seed super-pixel and object module first
(5.4.2) chooses the super-pixel s closest to object module1As start node, infer maximum spanning tree simultaneously with this
Obtain corresponding optimal subtree and its corresponding segmentation resultJudge the similarity degree of this segmentation result and object module: such as
Fruit similarity degree is eligibleThen think that segmentation result is effective, otherwise willIt is set as empty setAnd the wrong seed super-pixel for including in segmentation area is fed back toCarry out deletion update;
(5.4.3) traversal setFind out the segmentation result corresponding to optimal subtree before
It whether there is seed super-pixel s other than regionk, then repeat if it exists more than step obtain segmentation resultSimilarly carry out with
The similarity of object module judges and subsequent processing, updates segmentation resultWith seed super-pixel set;
(5.4.4) is completed to seed super-pixel setWhole traversals after, we obtain finally being directed to image Ii's
Segmentation resultWith updated seed super-pixel setAnd object module is completed according to these information
Update and seed super-pixel constraint information update, to make the estimation of model more close to change present in real scene
The seed super-pixel for changing situation and debug, then begins to iteration next time.
(6) iterative segmentation: the object module in step 4 is updated according to the segmentation result that step 5 obtains, according to step 5 institute
The method stated, then be split;
(7) step 6 is repeated, until final segmentation result no longer changes to arrive final segmentation result.
Further, in the step 2, super-pixel significance valueMeasurement specifically:
By conspicuousness detection technique, to the i-th width image IiObtain original Saliency maps φi, then calculate each super-pixel
For the conspicuousness mean value for all pixels point for including as its measurement, specific calculating is as follows:
WhereinI-th width image IiIn m-th of super-pixel RimAverage significance value,Indicate j-th of pixel
Significance value, area (Rim) it is super-pixel RimThe number of pixels for including.
Super-pixel repeatability value wimMeasurement, specifically:
Each super-pixel minimum value at a distance from all super-pixel in other each images is measured, N-1 most narrow spacings are obtained
From { d (Rim, Ik)}k≠i, further to N-1 minimum range { d (Rim, Ik)}k≠iIt is averaging, obtains average minimumIts
Middle distance metric d (Rim, Ik) distance weighted by the vector distance based on hsv color and the bag of words based on SIFT feature
It arrives, specific as follows:
Wherein cimAnd gimRespectively represent the i-th width image IiIn m-th of super-pixel RimHsv color characteristic vector and SIFT
Bag of words characteristic vector, ckm′And gkm′Respectively represent kth width image IkIn a super-pixel R of m 'km′Hsv color Characteristic Vectors
Amount and SIFT bag of words characteristic vector;
Super-pixel repeatability metric weights w is calculated by sigmoid formulaim:
Wherein μ and σ is the parameter for controlling the sigmoid functional form, μ=0.5, σ=0.1.
Further, step 6 specifically:
(6.1) according to newest segmentation result, foreground target model before updating is allowed to be more nearly mesh to be split
Mark;
(6.2) it according to updated object module, regenerates all possible candidate subtree collection and estimates maximum
Spanning tree;
(6.3) it according to updated object module and maximum spanning tree, is searched for again using dynamic programming techniques by subtree
Gather the forest constituted, obtains segmentation result;
(6.4) judge whether to meet cut-off condition, i.e., whether last segmentation result no longer changes.If satisfied, then iteration knot
Beam;If not satisfied, then repeating (6.1)-(6.3).
Implementation result:
According to above-mentioned steps, several pictures chosen in iCoseg database carry out Target Segmentation.Fig. 4, which is illustrated, to be selected from
The picture of iCoseg carries out Segmentation of Multi-target test.From fig. 4, it can be seen that the present invention is for target to be split, there are scales, appearance
In the case that the acute variations such as state, illumination and image include multiple targets, accurate object segmentation result can be still obtained.
Claims (2)
1. a kind of image multiple target constrained based on super-pixel and structuring cooperates with dividing method, which is characterized in that comprising following
Step:
(1) image pre-segmentation: for the image data set I={ I comprising common objective object1..., INIn every piece image
Ii, i=1,2......, N carry out over-segmentation processing, obtain super-pixel collection
(2) automatic target is found: the super-pixel collection based on each imageCount each super-pixelSignificance value
With repeated value wim, and calculate super-pixelEvaluation of estimate scoreim, Evaluation of estimate is small
In 0.6 × max (scorei) super-pixel be set as background, by evaluation of estimate be more than or equal to 0.6 × max (scorei) super-pixel
It is set as prospect;max(scorei) it is super-pixel collectionThe evaluation of estimate of the middle maximum super-pixel of evaluation of estimate;
(3) classifier learns: being found by automatic target by the super-pixel collection in training setIt is divided into foreground and background, for every
One super-pixel is described using the characteristic vector of 2004 following dimensions: (a) obtaining 800 Dimension Vector Quantization of Linear Prediction by k mean cluster
Hsv color indicates;(b) the SIFT bag of words that multiple dimensioned intensive sampling obtains, multiple dimensioned intensive sampling be 1200 dimension, respectively with
16,24,32 pixels are the image block multi-scale sampling on side, and the sampling interval is 3 pixels;(c) 4 binaryzation features, to describe
The contact situation of super-pixel and four boundaries of image;Based on features above, the support vector machines learning method of standard is utilized
To obtain preceding background class device;
(4) Target Modeling: being based on step (3) sorted information, is based on hsv color space to foreground target object and establishes target
Model ΨfWith background model Ψb;Super-pixel or super-pixel combination and target are calculated separately using Hellinger distance metric method
Similarity degree between modelSimilarity degree between super-pixel or super-pixel combination and background model
(5) based on the segmentation of super-pixel: utilizing object module ΨfWith background model Ψb, using the algorithm of Combinatorial Optimization to super picture
The subseries again of background before element carries out, to obtain the final segmentation of target object;It is proposed forest model it is assumed that assuming every
A super-pixel corresponds to a vertex, and single goal is divided, and last segmentation result is made of the super-pixel of multiple adjoinings, and
It can be expressed as adjacent mapSubtree;For Segmentation of Multi-target, last segmentation result is expressed as adjacent mapMultiple sons
Set the forest constituted;To sum up, it is assumed that last segmentation result is adjacent mapMultiple subtrees constitute forest;It is adjacent by establishing
Map interlinkingTo infer that the method for subtree collection determines last segmentation result;The specific implementation process is as follows:
(5.1) adjacent map is constructed: assuming that each super-pixel in image corresponds to a vertex in figure, two adjacent super pictures
It is connected between element by a line, thus constitutes adjacent mapFor final target object segmentation result,
Assuming that the forest that its multiple subtree for including by adjacent map is constituted;
(5.2) it establishes numerical model solution: establishing numerical model, the problem of Target Segmentation is converted into asking for combinatorial optimization problem
Solution is as follows:
When R is super-pixel or super-pixel combination in prospect,When R ' is the super-pixel or super-pixel group in background
When conjunction,Constraint condition indicates one kind before can only belonging to for any one super-pixel R in background;Pass through derivation
It obtains, to solution segmentation result, actually can be exchanged into the method for solving optimal subtree collection, and require optimal subtree set
It closes, needs first to estimate maximum spanning tree;
(5.3) it derives maximum spanning tree: obtaining all possible candidate subtree by the beam search method of beam search
SetBased on candidate subtree collectionMaximum spanning tree is obtained by the method for maximal possibility estimationIt derives as follows:
Indicate all potential spanning tree set,Indicate data likelihood probability, it is final to export,
For candidate subtree collection,For a certain subtree,Expression pairMaximum likelihood estimate
Meter, δ () are indicator function, δ ((x, y) ∈ Cq) indicate whether side (x, y) belongs to a certain subtree Cq; For
Subtree CqWith the similarity degree of object module, indicating whether side belongs to a certain subtree, P (x, y) indicates the probability of side (x, y),For the maximal possibility estimation to P (x, y);Maximum spanning tree is obtained by above formulaMaximal possibility estimation;
(5.4) search segmentation subtree collection: it is based on maximum spanning treeMaximal possibility estimation acquireThen pass through dynamic
Planning technology existsMiddle search obtains optimal subtree collection, and the specific implementation steps are as follows:
(5.4.1) is for image Ii, preceding background class device is acted on into super-pixel collectionObtain the super picture of seed for being classified as prospect
Element setIts seed super-pixel setBy discrete seed super-pixel structure
At being ranked up to obtain according to the similarity degree of each seed super-pixel and object module first
(5.4.2) chooses the super-pixel S closest to object module1As start node, maximum spanning tree is inferred with this and obtains phase
The optimal subtree answered and its corresponding segmentation resultJudge the similarity degree of this segmentation result to object module: if similar
Degree is eligible:Then think that segmentation result is effective, otherwise willIt is set as empty set:And the wrong seed super-pixel for including in segmentation area is fed back toCarry out deletion update;
(5.4.3) traversal setFind out the segmentation area corresponding to optimal subtree before
It whether there is seed super-pixel s in additionk, then repeat if it exists more than step obtain segmentation resultSimilarly progress and target
The similarity of model judges and subsequent processing, updates segmentation resultWith seed super-pixel set;
(5.4.4) is completed to seed super-pixel setWhole traversals after, we obtain finally being directed to image IiSegmentation
As a resultWith updated seed super-pixel setAnd object module is completed more according to these information
New and seed super-pixel constraint information update, to make the estimation of model more close to changing feelings present in real scene
Condition and the seed super-pixel of debug, then begin to iteration next time;
(6) iterative segmentation: the object module in step 4 is updated according to the segmentation result that step 5 obtains, according to described in step 5
Method, then be split;
(7) step 6 is repeated, until final segmentation result no longer changes to arrive final segmentation result.
2. according to the method described in claim 1, it is characterized by: step 6 specifically:
(6.1) according to newest segmentation result, foreground target model before updating is allowed to be more nearly target to be split;
(6.2) it according to updated object module, regenerates all possible candidate subtree collection and estimates maximum generation
Tree;
(6.3) it according to updated object module and maximum spanning tree, is searched for again using dynamic programming techniques by subtree collection
The forest of composition, obtains segmentation result;
(6.4) judge whether to meet cut-off condition, i.e., whether last segmentation result no longer changes;If satisfied, then iteration terminates;If
It is unsatisfactory for, then repeats (6.1)-(6.3).
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610120407.7A CN105809672B (en) | 2016-03-03 | 2016-03-03 | A kind of image multiple target collaboration dividing method constrained based on super-pixel and structuring |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610120407.7A CN105809672B (en) | 2016-03-03 | 2016-03-03 | A kind of image multiple target collaboration dividing method constrained based on super-pixel and structuring |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105809672A CN105809672A (en) | 2016-07-27 |
CN105809672B true CN105809672B (en) | 2019-09-13 |
Family
ID=56466027
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610120407.7A Expired - Fee Related CN105809672B (en) | 2016-03-03 | 2016-03-03 | A kind of image multiple target collaboration dividing method constrained based on super-pixel and structuring |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105809672B (en) |
Families Citing this family (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106875406B (en) | 2017-01-24 | 2020-04-14 | 北京航空航天大学 | Image-guided video semantic object segmentation method and device |
CN107103326B (en) * | 2017-04-26 | 2020-06-02 | 苏州大学 | Collaborative significance detection method based on super-pixel clustering |
CN107256412B (en) * | 2017-05-26 | 2019-07-12 | 东南大学 | A kind of figure building method based on more human eye perceptual grouping characteristics |
CN107527348B (en) * | 2017-07-11 | 2020-10-30 | 湖州师范学院 | Significance detection method based on multi-scale segmentation |
CN107610133B (en) * | 2017-08-28 | 2020-08-25 | 昆明理工大学 | Multi-target garment image collaborative segmentation method |
CN107909079B (en) * | 2017-10-11 | 2021-06-04 | 天津大学 | Cooperative significance detection method |
CN107909576B (en) * | 2017-11-22 | 2021-06-25 | 南开大学 | Indoor RGB-D image object segmentation method based on support semantic relation |
CN108305258B (en) * | 2018-01-31 | 2022-07-26 | 成都快眼科技有限公司 | Super-pixel segmentation method, system and storage device based on minimum spanning tree |
CN111223118A (en) * | 2018-11-27 | 2020-06-02 | 富士通株式会社 | Image processing apparatus, image processing method, and computer-readable recording medium |
CN109559321A (en) * | 2018-11-28 | 2019-04-02 | 清华大学 | A kind of sonar image dividing method and equipment |
CN109934235B (en) * | 2019-03-20 | 2021-04-20 | 中南大学 | Unsupervised abdominal CT sequence image multi-organ simultaneous automatic segmentation method |
CN110598711B (en) * | 2019-08-31 | 2022-12-16 | 华南理工大学 | Target segmentation method combined with classification task |
CN112164087B (en) * | 2020-10-13 | 2023-12-08 | 北京无线电测量研究所 | Super-pixel segmentation method and device based on edge constraint and segmentation boundary search |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101756696A (en) * | 2009-12-31 | 2010-06-30 | 中国人民解放军空军总医院 | Multiphoton skin lens image automatic analytical system and method for diagnosing malignant melanoma by using same system |
CN105046714A (en) * | 2015-08-18 | 2015-11-11 | 浙江大学 | Unsupervised image segmentation method based on super pixels and target discovering mechanism |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2014149563A (en) * | 2013-01-31 | 2014-08-21 | Akita Univ | Frame division device and frame division program |
-
2016
- 2016-03-03 CN CN201610120407.7A patent/CN105809672B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101756696A (en) * | 2009-12-31 | 2010-06-30 | 中国人民解放军空军总医院 | Multiphoton skin lens image automatic analytical system and method for diagnosing malignant melanoma by using same system |
CN105046714A (en) * | 2015-08-18 | 2015-11-11 | 浙江大学 | Unsupervised image segmentation method based on super pixels and target discovering mechanism |
Also Published As
Publication number | Publication date |
---|---|
CN105809672A (en) | 2016-07-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105809672B (en) | A kind of image multiple target collaboration dividing method constrained based on super-pixel and structuring | |
Ibrahim et al. | Image segmentation methods based on superpixel techniques: A survey | |
CN109948425B (en) | Pedestrian searching method and device for structure-aware self-attention and online instance aggregation matching | |
CN108960140B (en) | Pedestrian re-identification method based on multi-region feature extraction and fusion | |
CN109190524B (en) | Human body action recognition method based on generation of confrontation network | |
CN106055576B (en) | A kind of fast and effectively image search method under large-scale data background | |
CN107330451B (en) | Clothing attribute retrieval method based on deep convolutional neural network | |
EP2811424B1 (en) | Method and apparatus for training an estimator for estimating a pose of an articulated object | |
CN109165540B (en) | Pedestrian searching method and device based on prior candidate box selection strategy | |
CN108288051B (en) | Pedestrian re-recognition model training method and device, electronic equipment and storage medium | |
CN105740915B (en) | A kind of collaboration dividing method merging perception information | |
CN108319957A (en) | A kind of large-scale point cloud semantic segmentation method based on overtrick figure | |
CN107256017B (en) | Route planning method and system | |
CN106845430A (en) | Pedestrian detection and tracking based on acceleration region convolutional neural networks | |
CN106570480B (en) | A kind of human action classification method based on gesture recognition | |
CN114841257B (en) | Small sample target detection method based on self-supervision comparison constraint | |
CN105243139A (en) | Deep learning based three-dimensional model retrieval method and retrieval device thereof | |
CN112132014B (en) | Target re-identification method and system based on non-supervised pyramid similarity learning | |
CN105046714A (en) | Unsupervised image segmentation method based on super pixels and target discovering mechanism | |
CN112884742A (en) | Multi-algorithm fusion-based multi-target real-time detection, identification and tracking method | |
CN108846404A (en) | A kind of image significance detection method and device based on the sequence of related constraint figure | |
CN110147841A (en) | The fine grit classification method for being detected and being divided based on Weakly supervised and unsupervised component | |
CN112733602B (en) | Relation-guided pedestrian attribute identification method | |
CN110163130B (en) | Feature pre-alignment random forest classification system and method for gesture recognition | |
CN113032613B (en) | Three-dimensional model retrieval method based on interactive attention convolution neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190913 |