CN109165565A - A kind of video object discovery and dividing method based on Coupled Dynamic Markov Network - Google Patents
A kind of video object discovery and dividing method based on Coupled Dynamic Markov Network Download PDFInfo
- Publication number
- CN109165565A CN109165565A CN201810865881.1A CN201810865881A CN109165565A CN 109165565 A CN109165565 A CN 109165565A CN 201810865881 A CN201810865881 A CN 201810865881A CN 109165565 A CN109165565 A CN 109165565A
- Authority
- CN
- China
- Prior art keywords
- target
- frame
- video
- indicate
- segmentation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
Abstract
The video object discovery and dividing method that the invention discloses a kind of based on Coupled Dynamic Markov Network, belong to computer vision and area of pattern recognition, step includes: to divide two problems with video object to video object discovery first with dynamic Markov Network to model；Then appearance and video timing information are utilized, likelihood function, compatibility function and the kinematic function in model are modeled, and target detection model is initialized；Overall model is finally solved in the way of confidence spread, obtains the video object discovery with maximum a posteriori probability and segmentation result.The present invention can solve the problems, such as video object discovery and segmentation two simultaneously, and can realize the combined optimization of two problems, finally the accuracy rate of target detection and Target Segmentation can be made all to be highly improved.
Description
Technical field
The invention belongs to computer visions and mode identification technology, in particular to a kind of to be based on Coupled Dynamic Ma Erke
The video object of husband's network is found and dividing method.
Background technique
Along with the rapid growth of video data, the significance level of the technology for automatically treating of video is growing day by day.Wherein regard
Frequency target detection and Target Segmentation problem have increasing theoretical research value and practical application value.How to find automatically
The target that includes in video and be located out in the time and space be still current video process field difficulties.It should
Practical problem includes two sub-problems, i.e. video object discovery is divided with video object.In the problem, input is comprising target
It also include the video sequence of background, output is divided into two parts, one is to the timi requirement of target, i.e. the result of target detection；
The second is the space orientation to target, the i.e. result of Target Segmentation.At present traditional method be by two sub-problems Independent modeling,
The association between subproblem is not accounted for, two sub-problems cannot obtain combined optimization and solve.
Summary of the invention
The purpose of the present invention is to provide a kind of video object discovery and segmentation based on Coupled Dynamic Markov Network
Method, to solve above-mentioned technical problem.The present invention is based on Coupled Dynamic Markov Network by unified model
Model, can by video object find with video object segmentation two sub-problems be unified under a frame, two sons can be asked
Topic carries out combined optimization and solves；It may make the solution of two sub-problems cooperates with to promote, can finally two sub-problems be made all to reach
Higher accuracy rate.
In order to achieve the above objectives, the invention adopts the following technical scheme:
A kind of video object discovery and dividing method based on Coupled Dynamic Markov Network, comprising the following steps:
Step 1, comprising the video sequence of targetWherein ftIndicate the t frame image of video, video sequence V
It altogether include T frame image；Video sequence V is modeled using dynamic Markov Network, obtains Coupled Dynamic Markov Network frame
Frame model；
In the Coupled Dynamic Markov Network frame model of acquisition including target detection dynamic Markov Network and
The dynamic Markov Network of Target Segmentation；The observation level of the dynamic Markov Network of target detection is that image object is candidate
Region；The observation level of the dynamic Markov Network of Target Segmentation is image superpixel；
Step 2, the video sequence in step 1 including target is utilizedBetween the appearance and video consecutive frame for including
Timing information, the observation likelihood function in Coupled Dynamic Markov Network frame model, the compatibility function that step 1 is obtained
It is modeled with consecutive frame kinematic function, obtains Coupled Dynamic Markov Network model；And to the Coupled Dynamic Ma Er of acquisition
Target detection model in section's husband's network model initializes, and obtains initial target discovery result；
Step 3: solving the Coupled Dynamic Markov Network model that institute's step 2 obtains using belief propagation algorithm, obtain
To the result of video object discovery and video object segmentation.
Further, include: to the video sequence V specific steps modeled using dynamic Markov Network in step 1
Step 1.1, the target detection label of video sequence V is denoted asCorresponding target detection is observed Wherein ot,iIndicate ftI-th of candidate region；
Step 1.2, the Target Segmentation label of video sequence V is denoted asCorresponding Target Segmentation is observed Wherein st,jIndicate ftJ-th of super-pixel；
Step 1.3,WithBetween there are compatibility function Ψ (L, B).
Further, in step 2, to the observation likelihood letter in the Coupled Dynamic Markov Network frame model of acquisition
Number, compatibility function and consecutive frame kinematic function modeled specifically includes the following steps:
Step 2.1, it gives a mark to each candidate region, then establishes the likelihood function p (O of target detectiont|Lt)；Each candidate
The score in region are as follows:
r(ot,i)=rs(ot,i)·ra(ot,i)·rm(ot,i)
Wherein, rs(ot,i) indicate conspicuousness score, for indicating the significance degree of corresponding region；ra(ot,i) indicate object
Score, for indicating that corresponding region includes the confidence level of object；rm(ot,i) indicate sports scores, for indicating that corresponding region is wrapped
The confidence level of the object containing persistent movement；In the score r (o for obtaining each candidate targett,i) after, the likelihood function of target detection
It is established as formula:
Wherein,Indicate rs(ot,i) score after normalized；
Step 2.2, the likelihood function p (S about Target Segmentation is establishedt|Bt), concrete mode are as follows: the prospect for learning video is high
This mixed model and background gauss hybrid models, wherein prospect Gauss model is expressed as h1；Background Gauss model is h0；Target point
The likelihood function cut is established as formula:
Wherein,
Step 2.3, the foundation of compatibility function, will be from LtTo BtCompatibility function is defined as:
ΨLB(Lt,Bt)=IoU (ot,i,Bt(1))；i∈{1,…,K}
Wherein, Bt(1) segmentation result calculated, o are indicatedt,iIndicate ftI-th of candidate region；
It will be from BtTo LtCompatibility function is defined as:
Wherein, Ot(1) indicate the target detection calculated as a result, st,jIndicate ftJ-th of super-pixel；
Step 2.4, by the dynamic model p (L of target detectiont|Lt-1) is defined as:
Wherein,Indicate candidate region ot,iFor its time adjacent area ot-1,mTransition probability, i indicate present frame be
The label of candidate region in t frame, m indicate to be chosen as o in the former frame i.e. t-1 frame of t framet,iAdjacent area mark
Number；
By the dynamic model of Target Segmentation is defined as:
Wherein,Indicate super-pixel st,jFor its time neighbouring super pixels st-1,nTransition probability, j indicate present frame be
The label of super-pixel in t frame, n indicate to be chosen as s in the former frame i.e. t-1 frame of t framet,jAdjacent area label.
Further, in step 2, target detection model initialization specifically includes the following steps:
(1) video frame for having been illustrated as feature vector is divided by two classes by classifier, positive sample is equal from video
The frame of even acquisition, negative sample are the image in uncorrelated data set；
(2) classifier is trained by the data in step (1), and the confidence level for belonging to positive sample is greater than 80%
Sample be considered as positive sample, the confidence level for belonging to positive sample is considered as negative sample less than 30% sample；
(3) step (2) are repeated until classifier is restrained；
(4) classifier obtained using step (3) carries out two classification to all frames of video sequence V, obtains comprising target
Frame is with the preliminary classification for not including target frame as a result, realizing the initialization of target detection model.
Further, the process for being solved established model in step 3 using belief propagation algorithm, is specifically included following
Step:
(1) confidence level for considering adjacent before and after frames is calculated, wherein the confidence level from B to L and the confidence level difference from L to B
Are as follows:
WhereinFor the target observation since video to t moment；For from video
Terminate the target observation to t moment；Start to the segmentation of t moment to observe for video；For
Terminate the target observation to t moment from video；
(2) it after obtaining confidence level by step 1, is sent out according to the target that belief propagation algorithm obtains maximum a posteriori probability
Now result p (Lt| O, S) are as follows:
Obtain the Target Segmentation p (B of maximum a posteriori probabilityt| O, S) result are as follows:
Compared with prior art, the invention has the following advantages:
Video object based on Coupled Dynamic Markov Network of the invention is found and dividing method, utilizes dynamic Ma Er
Section's husband's network finds that dividing two problems with video object models to video object；Then by two dynamic decomposable markov networks
Network is coupled together, and overall model is solved in the way of confidence spread, and obtaining, there is the video object of maximum a posteriori probability to send out
Now with segmentation result.The present invention can solve the problems, such as video object discovery and segmentation two simultaneously, and the present invention can be real
The combined optimization of existing two problems has mathematically obtained having maximum a posteriori under target detection and Target Segmentation double-run
Target detection result and the object segmentation result (i.e. p (L of probabilityt| O, S) and p (Bt| O, S)), it finally may make target detection
It is all highly improved with the accuracy rate of Target Segmentation.
Further, the present invention is when initialized target finds model, using a kind of unsupervised repetitive exercise method.
This method does not depend on the data manually marked.If method depends on the training of handmarking's data, identical method
Corresponding handmarking's data mating must be provided when handling different data, and flag data is one and expends a large amount of manpowers
With the work of time.This method does not depend on the data manually marked, is conducive to popularization and application of the invention.
Detailed description of the invention
Fig. 1 is a kind of stream of video object discovery and dividing method based on Coupled Dynamic Markov Network of the invention
Journey schematic block diagram；
Fig. 2 is that a kind of video object discovery based on Coupled Dynamic Markov Network of the invention is established with dividing method
Coupled Dynamic Markov Network model schematic block diagram；
Fig. 3 is first in a kind of video object discovery and dividing method based on Coupled Dynamic Markov Network of the invention
The schematic process flow diagram of beginningization target detection model；
Fig. 4 is the Comparative result schematic diagram that distinct methods evaluate and test Target Segmentation on SegTrack data set；
Fig. 5 is that the Comparative result that distinct methods evaluate and test target detection on Noisy-ViDiSeg data set is illustrated
Figure；
Fig. 6 is that the Comparative result that distinct methods evaluate and test Target Segmentation on Noisy-ViDiSeg data set is illustrated
Figure.
Specific embodiment
In order to illustrate more clearly of the present invention, the present invention will be further described with reference to the accompanying drawing.
Referring to Fig. 1 to Fig. 3, a kind of video object discovery and segmentation based on Coupled Dynamic Markov Network of the invention
Method, specifically includes the following steps:
Step 1: establishing Coupled Dynamic Markov Network.Provide the video sequence comprising simple targetWherein ftIndicate the t frame image of video.V is modeled using dynamic Markov Network, obtains Coupled Dynamic horse
Er Kefu network frame model, including the dynamic Markov Network of corresponding video object discovery, observation level is image
Object candidate area, and the dynamic Markov Network of corresponding video object segmentation, observation level are image superpixel.
Specific modeling procedure includes: in step 1
(1) for the video sequence V comprising simple target, target detection is labeled asIt is corresponding to be observed Wherein ot,iIndicate ftI-th of candidate region.
(2) for V, Target Segmentation is labeled asIt is corresponding to be observed Its
Middle st,jIndicate ftJ-th of super-pixel.
(3)WithBetween there are compatibility function Ψ (L, B).
(4) timing information is added in the Coupled Dynamic Markov Network of foundation.We define since video to t
The image object at moment is observedTerminate to the image object of t moment to be observed from videoSimilar, the image segmentation since video to t moment is observedFrom video knot
The image object of beam to t moment is observed
Step 2: timing information between video appearance and video consecutive frame is utilized, in Coupled Dynamic Markov Network
Likelihood function, compatibility function and kinematic function are modeled, and are initialized to target detection model.The model of foundation is joined
See Fig. 2, include section Points And lines in Fig. 2 model, step 1 is used to explain the physical meaning of node, and step 2 is for explaining line
Meaning and specific method for building up.
Specifically includes the following steps:
(1) for the likelihood function p (O of target detectiont|Lt) need to give a mark to each candidate target, mode of giving a mark are as follows:
r(ot,i)=rs(ot,i)·ra(ot,i)·rm(ot,i)
Wherein rs(ot,i) indicate conspicuousness score, that is, indicate that the significance degree of corresponding region, circular are benefit
Saliency maps are obtained with existing conspicuousness method, then calculate the average value of the conspicuousness of corresponding region；ra(ot,i) expression thing
The confidence level that body score, i.e. expression corresponding region include object, circular are the complete edge for calculating corresponding region
The ratio at shared all edges；rm(ot,i) indicate sports scores, i.e. confidence of the expression corresponding region comprising persistent movement object
Degree, circular and ra(ot,i) similar, image border is only replaced with into light stream edge, remaining is identical.Obtaining r
(ot,i) after, the likelihood function of target detection is established as formula:
WhereinIndicate the score after all scores normalize in all video frames.
(2) for the likelihood function p (S of Target Segmentationt|Bt), it needs to learn two Gaussian Mixtures of foreground and background to video
Model, the model for prospect are h1, the model for background is h0.The likelihood function of Target Segmentation is established as formula:
(3) for compatibility function establish us will be from LtTo BtCompatibility function is defined as:
ΨLB(Lt,Bt)=IoU (ot,i,Bt(1))；i∈{1,…,K}
Wherein Bt(1) segmentation result calculated is indicated.Similar, we will be from BtTo LtCompatibility function definition
Are as follows:
Wherein Ot(1) the target detection result calculated is indicated.
(4) for the dynamic model p (L of target detectiont|Lt-1) we is defined as:
Wherein,Indicate candidate region ot,iFor its time adjacent area ot-1,mTransition probability, i indicate present frame
The label of candidate region in (i.e. t frame), m indicate to be chosen as o in the former frame (i.e. t-1 frame) of t framet,iAdjacent region
The label in domain.Specific calculation are as follows:
M=argmaxi′∈{1,…,K}IoU(ot,i,Warp(ot-1,i′))
Wherein Warp (ot-1,i′) it is ft-1In region ot-1,i′F is transformed into according to light streamtIn corresponding region, δi=δ
(lt-1,m,lt,i) it is an indicator variable.It is in lt-1,m≠lt,iWhen be 1, be otherwise 0.αi=EMD (hc(ot-1,m),hc(ot,i))
Indicate ot-1,mWith ot,iColor histogram EMD distance.Indicate ot-1,iWith ot,mHistogram of gradients
Chi-Square measure.
Similar, we define the dynamic model of Target Segmentation are as follows:
Wherein,Indicate super-pixel st,jFor its time neighbouring super pixels st-1,nTransition probability, j indicate present frame (i.e.
T frame) in super-pixel label, n indicate t frame former frame (i.e. t-1 frame) in be chosen as st,jAdjacent area mark
Number.Specific calculation are as follows:
N=argmaxj'∈{1,…,J}IoU(st,j,Warp(st-1,j'))
Wherein Warp (st-1,j') it is ft-1In region st-1,j'F is transformed into according to light streamtIn corresponding region, δj=δ
(bt-1,n,bt,j) it is an indicator variable.It is in bt-1,n≠bt,jWhen be 1, be otherwise 0.ωj=| | hm(st-1,n)-hm(st,j)|
|2Indicate st-1,jWith st,nOptical flow gradient histogram Euclidean distance.σjIt is also an indicator variable st,jWith st-1,nBelong to mesh
It is 1 when the prospect that mark discovery obtains, is otherwise 0.μj=IoU (st,j,Warp(st-1,n)) it is st,jWith Warp (st-1,n) IoU point
Number.
The initialization process of step 2 target detection model is as shown in Figure 3.Specifically includes the following steps:
(1) video frame for having been illustrated as feature vector is divided into two classes using classifier, positive sample is equal from video
The frame of even acquisition, wherein may include noise, negative sample be the image in uncorrelated data set.
(2) classifier is trained using these data, and by and by belonging to the confidence level of positive sample greater than 80%
Sample is considered as positive sample, and the confidence level for belonging to positive sample is considered as negative sample less than 30% sample, continues to train.(3) it is straight that (2) are repeated
It is restrained to classifier, two classification, the available frame comprising target is carried out to all frames of video using finally obtained classifier
With the preliminary classification not comprising target frame as a result, the initialization result of as target detection.
Step 3: established model is solved using belief propagation algorithm.Specifically includes the following steps:
(1) certainty factor algebra is generalized in time series, wherein the confidence level m from B to LBL(Lt) and setting from L to B
Reliability mLB(Bt) be respectively as follows:
After obtaining confidence level, according to the target detection result of the available maximum a posteriori probability of belief propagation algorithm
p(Lt| O, S) are as follows:
Similar, the Target Segmentation p (B of maximum a posteriori probabilityt| O, S) result are as follows:
It is finally obtained the result is that posterior probability (i.e. p (L under two kinds of observation conditionst| O, S) and p (Bt| O, S)),
This mathematically shows our final results, unifies the discovery of combined optimization video object and asks with video object two sons of segmentation
Topic.In above-described embodiment,The video sequence comprising simple target is indicated, wherein wherein ftIndicate the t frame of video
Image, it includes T frame image that video, which has altogether,；Indicate the target detection label of video；Indicate video
Target detection observation, whereinot,iIndicate ftI-th of candidate region；Indicate the target of video
Dividing mark；Indicate the Target Segmentation observation of video, whereinst,jIndicate ftJ-th surpass picture
Element；Ψ (L, B): it indicatesWithBetween there are compatibility functions, particularly, LtTo BtCompatibility letter
Number is ΨLB(Lt,Bt), from BtTo LtCompatibility function be ΨBL(Bt,Lt)；p(Ot|Lt): indicate the observation of video object discovery
Likelihood function；p(St|Bt): indicate the observation likelihood function of video object segmentation；p(Lt|Lt-1): indicate video object discovery
Dynamic model；p(Bt|Bt-1): indicate the dynamic model of video object segmentation.
The present invention is suitable for automatically processing video, finds the main target in video and comes out it from background segment.This
The video object based on Coupled Dynamic Markov Network of invention is found and dividing method, first with dynamic decomposable markov networks
Network finds that dividing two problems with video object models to video object；Then appearance is utilized, the information such as video timing are right
Likelihood function, compatibility function and kinematic function in model are modeled, and are initialized to target detection model；Finally
Overall model is solved in the way of confidence spread, obtaining, there is the video object discovery of maximum a posteriori probability to tie with segmentation
Fruit.The present invention can solve the problems, such as video object discovery and segmentation two simultaneously, and two problems may be implemented in the present invention
Combined optimization, finally make the accuracy rate of target detection and Target Segmentation all be highly improved.
Data analysis
Fig. 4 be different method for evaluate video object segmentation public data collection SegTrack on experimental result,
As seen from the figure, in Target Segmentation, our method has all obtained most in addition to " monkey " and " soldier " video
It is high as a result, and comprehensively consider, that is, compare average result, achieve highest accuracy.
Fig. 5 be on Noisy-ViDiSeg data set to target detection evaluated and tested as a result, as seen from the figure,
Our method has all obtained highest accuracy rate in all videos in target detection.
Fig. 6 be on Noisy-ViDiSeg data set to Target Segmentation evaluated and tested as a result, as seen from the figure,
Our method has obtained highest accuracy rate in relatively more comprehensive all data average results in Target Segmentation.
In conclusion being compared with other methods compared with, our method in two problems of Target Segmentation and target detection all
Obtain biggish promotion.
Claims (5)
1. it is a kind of based on Coupled Dynamic Markov Network video object discovery and dividing method, which is characterized in that including with
Lower step:
Step 1, comprising the video sequence of targetWherein ftIndicate that the t frame image of video, video sequence V wrap altogether
The image of frame containing T；Video sequence V is modeled using dynamic Markov Network, obtains Coupled Dynamic Markov Network frame mould
Type；
It include the dynamic Markov Network and target of target detection in the Coupled Dynamic Markov Network frame model of acquisition
The dynamic Markov Network of segmentation；The observation level of the dynamic Markov Network of target detection is image object candidate regions
Domain；The observation level of the dynamic Markov Network of Target Segmentation is image superpixel；
Step 2, the video sequence in step 1 including target is utilizedTiming between the appearance and video consecutive frame for including
Information, observation likelihood function, compatibility function and the phase in Coupled Dynamic Markov Network frame model that step 1 is obtained
Adjacent frame kinematic function is modeled, and Coupled Dynamic Markov Network model is obtained；And to the Coupled Dynamic Markov of acquisition
Target detection model in network model is initialized, and initial target discovery result is obtained；
Step 3: the Coupled Dynamic Markov Network model that institute's step 2 obtains is solved using belief propagation algorithm, depending on
The result of frequency target detection and video object segmentation.
2. a kind of video object discovery and segmentation side based on Coupled Dynamic Markov Network according to claim 1
Method, which is characterized in that include: to the video sequence V specific steps modeled using dynamic Markov Network in step 1
Step 1.1, the target detection label of video sequence V is denoted asCorresponding target detection observation is denoted asWherein ot,iIndicate ftI-th of candidate region；
Step 1.2, the Target Segmentation label of video sequence V is denoted asCorresponding Target Segmentation observation is denoted asWherein st,jIndicate ftJ-th of super-pixel；
Step 1.3,WithBetween there are compatibility function Ψ (L, B).
3. a kind of video object discovery and segmentation side based on Coupled Dynamic Markov Network according to claim 2
Method, which is characterized in that in step 2, in the Coupled Dynamic Markov Network frame model of acquisition observation likelihood function,
Compatibility function and consecutive frame kinematic function modeled specifically includes the following steps:
Step 2.1, it gives a mark to each candidate region, then establishes the likelihood function p (O of target detectiont|Lt)；Each candidate region
Score are as follows:
r(ot,i)=rs(ot,i)·ra(ot,i)·rm(ot,i)
Wherein, rs(ot,i) indicate conspicuousness score, for indicating the significance degree of corresponding region；ra(ot,i) indicate object score,
For indicating that corresponding region includes the confidence level of object；rm(ot,i) indicate sports scores, for indicating that corresponding region includes to continue
The confidence level of moving object；In the score r (o for obtaining each candidate targett,i) after, the likelihood function of target detection is established as
Formula:
Wherein,Indicate rs(ot,i) score after normalized；
Step 2.2, the likelihood function p (S about Target Segmentation is establishedt|Bt), concrete mode are as follows: the prospect Gauss for learning video is mixed
Molding type and background gauss hybrid models, wherein prospect Gauss model is expressed as h1；Background Gauss model is h0；Target Segmentation
Likelihood function is established as formula:
Wherein,
Step 2.3, the foundation of compatibility function, will be from LtTo BtCompatibility function is defined as:
ΨLB(Lt,Bt)=IoU (ot,i,Bt(1))；i∈{1,…,K}
Wherein, Bt(1) segmentation result calculated, o are indicatedt,iIndicate ftI-th of candidate region；
It will be from BtTo LtCompatibility function is defined as:
Wherein, Ot(1) indicate the target detection calculated as a result, st,jIndicate ftJ-th of super-pixel；
Step 2.4, by the dynamic model p (L of target detectiont|Lt-1) is defined as:
Wherein,Indicate candidate region ot,iFor its time adjacent area ot-1,mTransition probability, i indicate present frame i.e. t frame
The label of middle candidate region, m indicate to be chosen as o in the former frame i.e. t-1 frame of t framet,iAdjacent area label；
By the dynamic model of Target Segmentation is defined as:
Wherein,Indicate super-pixel st,jFor its time neighbouring super pixels st-1,nTransition probability, j indicate present frame i.e. t frame
The label of middle super-pixel, n indicate to be chosen as s in the former frame i.e. t-1 frame of t framet,jAdjacent area label.
4. a kind of video object discovery and segmentation side based on Coupled Dynamic Markov Network according to claim 1
Method, which is characterized in that in step 2, the initialization of target detection model specifically includes the following steps:
(1) video frame for having been illustrated as feature vector is divided by two classes by classifier, positive sample is uniformly to adopt from video
The frame of collection, negative sample are the image in uncorrelated data set；
(2) classifier is trained by the data in step (1), and the confidence level for belonging to positive sample is greater than to 80% sample
Originally it is considered as positive sample, the confidence level for belonging to positive sample is considered as negative sample less than 30% sample；
(3) step (2) are repeated until classifier is restrained；
(4) classifier obtained using step (3) carries out two classification to all frames of video sequence V, obtain the frame comprising target with
Preliminary classification not comprising target frame is as a result, realize the initialization of target detection model.
5. a kind of video object discovery and segmentation side based on Coupled Dynamic Markov Network according to claim 4
Method, which is characterized in that the process for solving established model in step 3 using belief propagation algorithm specifically includes following step
It is rapid:
(1) confidence level for considering adjacent before and after frames is calculated, wherein the confidence level from B to L and the confidence level from L to B are respectively as follows:
WhereinFor the target observation since video to t moment；To terminate from video
To the target observation of t moment；Start to the segmentation of t moment to observe for video；For from view
Frequency terminates the target observation to t moment；
(2) after obtaining confidence level by step 1, the target detection knot of maximum a posteriori probability is obtained according to belief propagation algorithm
Fruit p (Lt| O, S) are as follows:
Obtain the Target Segmentation p (B of maximum a posteriori probabilityt| O, S) result are as follows:
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810865881.1A CN109165565A (en) | 2018-08-01 | 2018-08-01 | A kind of video object discovery and dividing method based on Coupled Dynamic Markov Network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810865881.1A CN109165565A (en) | 2018-08-01 | 2018-08-01 | A kind of video object discovery and dividing method based on Coupled Dynamic Markov Network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109165565A true CN109165565A (en) | 2019-01-08 |
Family
ID=64898648
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810865881.1A Pending CN109165565A (en) | 2018-08-01 | 2018-08-01 | A kind of video object discovery and dividing method based on Coupled Dynamic Markov Network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109165565A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110222649A (en) * | 2019-06-10 | 2019-09-10 | 北京达佳互联信息技术有限公司 | Video classification methods, device, electronic equipment and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100231802A1 (en) * | 2009-03-12 | 2010-09-16 | Sony Corporation | Method and system for carrying out reliability classification for motion vectors in a video |
| CN101840579A (en) * | 2010-01-28 | 2010-09-22 | 浙江大学 | Method for realizing multi-target tracking by using video segmentation and particle filter |
| CN102156995A (en) * | 2011-04-21 | 2011-08-17 | 北京理工大学 | Video movement foreground dividing method in moving camera |
| CN105869178A (en) * | 2016-04-26 | 2016-08-17 | 昆明理工大学 | Method for unsupervised segmentation of complex targets from dynamic scene based on multi-scale combination feature convex optimization |
| CN106296728A (en) * | 2016-07-27 | 2017-01-04 | 昆明理工大学 | A kind of Segmentation of Moving Object method in unrestricted scene based on full convolutional network |
-
2018
- 2018-08-01 CN CN201810865881.1A patent/CN109165565A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100231802A1 (en) * | 2009-03-12 | 2010-09-16 | Sony Corporation | Method and system for carrying out reliability classification for motion vectors in a video |
| CN101840579A (en) * | 2010-01-28 | 2010-09-22 | 浙江大学 | Method for realizing multi-target tracking by using video segmentation and particle filter |
| CN102156995A (en) * | 2011-04-21 | 2011-08-17 | 北京理工大学 | Video movement foreground dividing method in moving camera |
| CN105869178A (en) * | 2016-04-26 | 2016-08-17 | 昆明理工大学 | Method for unsupervised segmentation of complex targets from dynamic scene based on multi-scale combination feature convex optimization |
| CN106296728A (en) * | 2016-07-27 | 2017-01-04 | 昆明理工大学 | A kind of Segmentation of Moving Object method in unrestricted scene based on full convolutional network |
Non-Patent Citations (1)
| Title |
|---|
| ZIYI LIU 等: "Joint Video Object Discovery and Segmentation by Coupled Dynamic Markov Networks", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110222649A (en) * | 2019-06-10 | 2019-09-10 | 北京达佳互联信息技术有限公司 | Video classification methods, device, electronic equipment and storage medium |
| CN110222649B (en) * | 2019-06-10 | 2020-12-18 | 北京达佳互联信息技术有限公司 | Video classification method and device, electronic equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106997597B (en) | It is a kind of based on have supervision conspicuousness detection method for tracking target | |
| CN109829443B (en) | Video behavior identification method based on image enhancement and 3D convolution neural network | |
| CN109543606B (en) | Human face recognition method with attention mechanism | |
| CN104599275B (en) | The RGB-D scene understanding methods of imparametrization based on probability graph model | |
| CN104050471B (en) | Natural scene character detection method and system | |
| CN110276264B (en) | Crowd density estimation method based on foreground segmentation graph | |
| CN106778796B (en) | Human body action recognition method and system based on hybrid cooperative training | |
| CN103942751B (en) | A kind of video key frame extracting method | |
| CN110111340A (en) | The Weakly supervised example dividing method cut based on multichannel | |
| CN109949316A (en) | A kind of Weakly supervised example dividing method of grid equipment image based on RGB-T fusion | |
| CN106295584A (en) | Depth migration study is in the recognition methods of crowd's attribute | |
| CN109657612B (en) | Quality sorting system based on facial image features and application method thereof | |
| CN105205475A (en) | Dynamic gesture recognition method | |
| CN110827312B (en) | Learning method based on cooperative visual attention neural network | |
| CN105931253A (en) | Image segmentation method combined with semi-supervised learning | |
| CN107527054B (en) | Automatic foreground extraction method based on multi-view fusion | |
| CN111027377B (en) | Double-flow neural network time sequence action positioning method | |
| CN106504255A (en) | A kind of multi-Target Image joint dividing method based on multi-tag multi-instance learning | |
| CN108021869A (en) | A kind of convolutional neural networks tracking of combination gaussian kernel function | |
| CN106228109A (en) | A kind of action identification method based on skeleton motion track | |
| CN112364791B (en) | Pedestrian re-identification method and system based on generation of confrontation network | |
| CN112052772A (en) | Face shielding detection algorithm | |
| CN110956158A (en) | Pedestrian shielding re-identification method based on teacher and student learning frame | |
| CN106570885A (en) | Background modeling method based on brightness and texture fusion threshold value | |
| CN104680193A (en) | Online target classification method and system based on fast similarity network fusion algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190108 |