CN108710761A - A kind of robust Model approximating method removing outlier based on spectral clustering - Google Patents
A kind of robust Model approximating method removing outlier based on spectral clustering Download PDFInfo
- Publication number
- CN108710761A CN108710761A CN201810494460.2A CN201810494460A CN108710761A CN 108710761 A CN108710761 A CN 108710761A CN 201810494460 A CN201810494460 A CN 201810494460A CN 108710761 A CN108710761 A CN 108710761A
- Authority
- CN
- China
- Prior art keywords
- model
- outlier
- classification
- spectral clustering
- interior
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2323—Non-hierarchical techniques based on graph theory, e.g. minimum spanning trees [MST] or graph cuts
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Geometry (AREA)
- Computer Hardware Design (AREA)
- Discrete Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of robust Model approximating methods removing outlier based on spectral clustering, generate model hypothesis by carrying out multiple repairing weld to input data, build the similar matrix between preference matrix and its row vector;Then using based on spectral clustering removal outlier and the more structural model data class of generation;Finally according to the stopping function of more structural model instance datas come judge simulate approximating method whether stop, obtaining final models fitting as a result, the model parameter of i.e. more structural model structures.The present invention furthers investigate distinction of outlier during spectral clustering with interior point, efficiently removes outlier, accurately estimates more structural model parameters, and interior point and the result of outlier classification can instruct the subsequent sampling process of local sampling strategy.
Description
Technical field
The present invention relates to computer vision model-fitting technique fields, and in particular to one kind removing outlier based on spectral clustering
Robust Model approximating method.
Background technology
The models fitting basic research task important as one is played an important role in field of machine vision, regarding
Feel that SLAM, motion segmentation, three-dimensional reconstruction, panorama are taken pictures etc. all to have a wide range of applications.As shown in Figure 1, models fitting side
The task of method is to generate model hypothesis by multiple data sampling, and the mould of these structure examples is then estimated by model selection
Shape parameter.Since the original image and video information obtained from sensor is by camera inherent parameters, shooting angle and distance, light
According to the influence of equal environmental changes, the key feature in image is often selected to describe sub (such as local feature description in models fitting
Sub- SIFT feature) it is used as input data.And noise (noise), outlier are inevitably present in these data
(outlier) information such as, accurately estimate model number and its corresponding parameter be still a great challenge task.
It is existing based on outlier removal pattern fitting method be usually required for independent step to containing outlier
It is removed after being differentiated, if removing outlier before model estimation, the interior point of some model instances may be removed simultaneously;Such as
Fruit removes after model estimation, and the presence of outlier can influence the accuracy of model parameter estimation;Secondly, most of method
Do not use the subsequent sampling process that local sampling strategy is instructed with the relevant information of model fully yet, lead to not for comprising
Noise and more structural model instance data quick samplings of high proportion outlier are to clean data subset.
Invention content
In view of the above-mentioned problems, the purpose of the present invention is to provide a kind of robust Model removing outlier based on spectral clustering is quasi-
Conjunction method can efficiently remove outlier, improve the accuracy of model parameter.
To achieve the above object, the technical solution adopted by the present invention is:
A kind of robust Model approximating method removing outlier based on spectral clustering, specifically includes following steps:
Similar matrix between step 1, structure preference matrix and its row vector;
From N number of observation data point X={ x1,x2,...,xNIn, M subsets of stochastical sampling, it is false that estimation obtains M model
If Θ={ θ1,θ2,...,θM};Then each model hypothesis is given to assign corresponding weight:
Wherein, nmIt is that m-th of the interior of model hypothesis is counted out;It is i-th of data point relative to m-th of model hypothesis
Residual error;SmIt is the interior spot noise scale estimated by IKOSE methods;ψ () and hmIt is kernel function and its corresponding bandwidth;
The adaptive weight threshold obtained by IKOSE methods removes the low model hypothesis of weight, retains G more
Add model hypothesis Θ={ θ of robust1,θ2,...,θG};So, preference of N number of observation data point relative to this G model hypothesis
Matrix is represented by:
Wherein, P is the two-dimensional matrix of N*M dimensions;ri gIt is residual error of i-th of data point relative to g-th of model hypothesis;
SgIt is interior spot noise scale;
By in preference matrix row vector P (i,:) and P (j,:) distance d (P (i,:),P(j,:)) structure similar matrix:
Wherein, σ is the parameter of exponential function;
Step 2, the subspace classification that similar matrix is obtained by spectral clustering, carry out according to the classification score of concept subspace
Outlier differentiates;
The Spectral Clustering that can automatically determine subspace number is used to obtain the k sub-spaces of similar matrix as { C1,
C2,...,Ck, the concept space of every sub-spaces classification of similar matrix is then built, and calculate the classification point of concept subspace
Number, wherein the label that each data point belongs to some structural model example is I={ i1,i2,...,iN};
Then the classification score of concept subspace is differentiated by following formula:
Wherein, n (Ck) indicate to belong to classification CkInterior point, as S (Ck) being more than certain value, then the category is interior classification, it is no
It is then outlier classification;Outlier classification is removed, the corresponding interior point of more structural model data class is retained;
Step 3 judges whether to stop models fitting according to stopping function;
First, the interior point of obtained model data class is fitted according to least square method, each model can be obtained
Parameter:
Wherein, θlIt is the model parameter that first of model structure is obtained by least square method;Expression passes through spectrum
Cluster all interior points of obtained l model data classes;
Then, the residual sum of squares (RSS) that the corresponding interior point of each model parameter obtains is counted:
Finally, we obtain the item whether models fitting stops according to the difference of the residual sum of squares (RSS) of front and back iteration twice
Part:
ft(θl)-ft-1(θl) < δ (8);
If difference is less than threshold value δ, iteration ends export the model of outlier classification and more structural model examples
Parameter;Otherwise, continue sampled point subset in the interior point data of more structural model data categories, estimate model hypothesis and build phase
Like matrix, then repeatedly step 2 and step 3 continue next iteration, until meeting formula (8) or reaching greatest iteration time
Number.
The kernel function is:
Wherein, t is the threshold constant of setting.
The maximum iteration is 5 times.
After adopting the above scheme, the present invention generates model hypothesis by carrying out multiple repairing weld to input data, builds preference
Similar matrix between matrix and its row vector;Then using based on spectral clustering removal outlier and the more structural models of generation
Data class;Finally judge to simulate whether approximating method stops according to the stopping function of more structural model instance datas, obtain most
Whole models fitting as a result, the model parameter of i.e. more structural model structures.The present invention furthers investigate outlier in spectral clustering mistake
The distinction put in Cheng Zhongyu, efficiently removes outlier, accurately estimates more structural model parameters, and interior point and outlier classification
Result can instruct the subsequent sampling process of local sampling strategy.
Description of the drawings
Fig. 1 is existing pattern fitting method schematic diagram;
Fig. 2 is flow chart of the method for the present invention;
Fig. 3 is the subspace classification range distribution schematic diagram obtained based on Spectral Clustering.
Specific implementation mode
As shown in Fig. 2, present invention is disclosed a kind of robust Model approximating method removing outlier based on spectral clustering, tool
Body includes the following steps:
Similar matrix between step 1, structure preference matrix and its row vector;
It is assumed that from N number of observation data point X={ x1,x2,...,xNIn, M subsets of stochastical sampling, estimation obtains M
Model hypothesis Θ={ θ1,θ2,...,θM, it is then quasi- in model in order to reduce some redundancies and poor robustness model hypothesis
Influence during conjunction, we assign corresponding weight to each model hypothesis:
Wherein, nmIt is that m-th of the interior of model hypothesis is counted out;ri mIt is i-th of data point relative to m-th of model hypothesis
Residual error;SmIt is the interior spot noise scale estimated by IKOSE methods;ψ () and hmIt is kernel function and its corresponding bandwidth.
The expression formula for the kernel function that the present invention uses is as follows:
Wherein, t is the threshold constant of setting, we set it to 2.5.
The adaptive weight threshold obtained again by IKOSE methods removes the lower model hypothesis of some weights,
Retain G more robust model hypothesis Θ={ θ1,θ2,...,θG}.So, N number of observation data point is relative to this G model
The preference matrix of hypothesis is represented by:
Wherein, P is the two-dimensional matrix of N*M dimensions;ri gIt is residual error of i-th of data point relative to g-th of model hypothesis;
SgIt is interior spot noise scale.
By in preference matrix row vector P (i,:) and P (j,:) distance d (P (i,:),P(j,:)) it may make up similar square
Battle array:
Wherein, σ is the parameter of exponential function.
Step 2, the subspace classification that similar matrix is obtained by spectral clustering, carry out according to the classification score of concept subspace
Outlier differentiates;
It is found in previous research:To row vector P (i, m) the structure concept space of preference matrix, can be obtained interior point and
The different range distribution of outlier can remove the outlier in data according to the size of range distribution.And this patent passes through experiment
It was found that it is that five lines are based on spectral clustering also to have similar property, Fig. 3 in concept space of the similar matrix per sub-spaces classification
The range distribution schematic diagram of data point of the subspace classification that method obtains in concept space in classification.
Therefore, the present invention is obtained similar using the Spectral Clustering (Self-Tuning) that can automatically determine subspace number
The k sub-spaces of matrix are { C1,C2,...,Ck, every sub-spaces classification of similar matrix is then built in concept space
Range distribution, and calculate the classification score of concept subspace, judge its category attribute (interior classification or outlier classification),
In each data point to belong to the label of some structural model example (classification) be I={ I1,I2,...,IN}.On this basis not only
The classification containing larger outlier ratio can be obtained, and can be to the interior point minute of the model instance comprising more structured datas
Class.The numerical score that the classification of outlier can be obtained by formula (5) is differentiated:
Wherein, wherein IiRefer to i-th of data and belongs to which subspace classification, each concept subspace classification CkIt will basis
All data points obtain its fractional value S (C in its classificationk)。n(Ck) indicate to belong to classification CkInterior point, as S (Ck) be more than centainly
Then the category is interior classification to value, is otherwise outlier classification.Outlier classification is removed, it is each to retain more structural model data class
Self-corresponding interior point.
Differentiate that outlier is advantageous in that in the subspace structure concept space of similar matrix:In the subspace classification of generation
In may include multiple outlier classes, multiple and different distributions is fitted to outlier in this way, is more nearly truthful data.
Step 3 judges whether to stop models fitting according to stopping function;
First, the interior point of obtained model data class is fitted according to least square method (Least Square), it can
To obtain the parameter of each model:
Wherein, θlIt is the model parameter that first of model structure is obtained by least square method;Expression passes through spectrum
Cluster all interior points of obtained l model data classes.
Then, the residual sum of squares (RSS) that the corresponding interior point of each model parameter obtains is counted:
Finally, we obtain the item whether models fitting stops according to the difference of the residual sum of squares (RSS) of front and back iteration twice
Part:
ft(θl)-ft-1(θl) < δ (8);
If difference is less than threshold value δ, iteration ends export the model of outlier classification and more structural model examples
Parameter;Otherwise, continue sampled point subset in the interior point data of more structural model data categories, estimate model hypothesis and build phase
Like matrix, then repeatedly step 2 and step 3 continue next iteration, until meeting formula (8) or reaching maximum iteration
(5 times).
When for more structural model examples of relatively simple lines the case where, pass through one to cluster process twice
Simultaneously accurately to export the model parameter of outlier and more structured datas.And when in face of needing to estimate that the list of more structural models is answered
It, can be according to once clustering as a result, change in obtained interior point data guidance next time when the complex situations such as matrix and basis matrix
For the sampling of process, generate more robust comprising more interior model hypothesis put.This have the advantage that by each iteration,
We can obtain cleaner and more robust model hypothesis, and preferably distinguish outlier information.
In order to verify the performance of the present invention, above-mentioned pattern fitting method, code fortune are realized with Matlab Programming with Pascal Language
Capable hardware platform is 8 core processors of 3.4GHZ.More structural models of five straight lines of the selection comprising high proportion outlier are made
For test data of experiment collection, including more structural models of five straight lines in the test set, it is 50 to count out in every straight line,
Peeling off, to count out be 250, and the total number of data point is 500.
Experiment 50 times is repeated to the test set, the point number of subsets of each initial random acquisition is 2000.We utilize 50 times
Average false drop rate and minimum false drop rate as evaluation criterion, while giving the run time of each algorithm as a comparison,
The calculation formula of middle false drop rate is as follows:
The pattern fitting method and the classical model approximating method based on outlier removing method that table 1 gives the present invention
Comparison result, it is specific as shown in table 1.
Method | Average false drop rate (%) | Minimum false drop rate (%) | Run time (second) |
KF | 25.02 | 16.6 | 2.59 |
T-linkage | 26.07 | 19.6 | 24.87 |
The pattern fitting method of the present invention | 16.02 | 12.6 | 1.94 |
Table 1
In table 1, KF be model estimation before remove outlier method, T-linkage be model estimation after remove from
The method of group's point.Result can be seen that the pattern fitting method of the present invention is substantially better than other methods from table, achieve minimum
Average false drop rate (16.02%) and minimum false drop rate (12.6%);Simultaneously at runtime on, this patent propose method
(1.94 seconds) embody the high efficiency of this patent method also below other control methods.
To sum up, the robust Model approximating method proposed by the present invention that outlier is removed based on spectral clustering can reach efficient standard
True effect, to provide preferably theoretical base for the practical application of more structural model approximating methods comprising high proportion outlier
Plinth.
The above is only the embodiment of the present invention, is not intended to limit the scope of the present invention, therefore every
According to the technical essence of the invention to any subtle modifications, equivalent variations and modifications made by above example, this is still fallen within
In the range of inventive technique scheme.
Claims (3)
1. a kind of robust Model approximating method removing outlier based on spectral clustering, it is characterised in that:The pattern fitting method
Specifically include following steps:
Similar matrix between step 1, structure preference matrix and its row vector;
From N number of observation data point X={ x1,x2,...,xNIn, M subsets of stochastical sampling, estimation obtains M model hypothesis Θ
={ θ1,θ2,...,θM};Then each model hypothesis is given to assign corresponding weight:
Wherein, nmIt is that m-th of the interior of model hypothesis is counted out;ri mIt is i-th of data point relative to the residual of m-th model hypothesis
Difference;SmIt is the interior spot noise scale estimated by IKOSE methods;ψ () and hmIt is kernel function and its corresponding bandwidth;
The adaptive weight threshold obtained by IKOSE methods removes the low model hypothesis of weight, retains G more Shandongs
Model hypothesis Θ={ θ of stick1,θ2,...,θG};So, preference matrix of N number of observation data point relative to this G model hypothesis
It is represented by:
Wherein, P is the two-dimensional matrix of N*M dimensions;ri gIt is residual error of i-th of data point relative to g-th of model hypothesis;SgIt is
Interior spot noise scale;
By in preference matrix row vector P (i,:) and P (j,:) distance d (P (i,:),P(j,:)) structure similar matrix:
Wherein, σ is the parameter of exponential function;
Step 2, the subspace classification that similar matrix is obtained by spectral clustering, peel off according to the classification score of concept subspace
Point differentiates;
The Spectral Clustering that can automatically determine subspace number is used to obtain the k sub-spaces of similar matrix as { C1,C2,...,
Ck, the concept space of every sub-spaces classification of similar matrix is then built, and calculate the classification score of concept subspace,
In each data point to belong to the label of some structural model example be I={ i1,i2,...,iN};
Then the classification score of concept subspace is differentiated by following formula:
Wherein, n (Ck) indicate to belong to classification CkInterior point, as S (Ck) be more than certain value then the category be interior classification, otherwise for
Outlier classification;Outlier classification is removed, the corresponding interior point of more structural model data class is retained;
Step 3 judges whether to stop models fitting according to stopping function;
First, the interior point of obtained model data class is fitted according to least square method, the ginseng of each model can be obtained
Number:
Wherein, θlIt is the model parameter that first of model structure is obtained by least square method;Expression passes through spectral clustering
All interior points of obtained l model data classes;
Then, the residual sum of squares (RSS) that the corresponding interior point of each model parameter obtains is counted:
Finally, we obtain the condition whether models fitting stops according to the difference of the residual sum of squares (RSS) of front and back iteration twice:
ft(θl)-ft-1(θl) < δ (8);
If difference is less than threshold value δ, iteration ends export outlier classification and the model ginseng of more structural model examples
Number;Otherwise, continue sampled point subset in the interior point data of more structural model data categories, estimate model hypothesis and build similar
Matrix, then repeatedly step 2 and step 3 continue next iteration, until meeting formula (8) or reaching maximum iteration.
2. a kind of robust Model approximating method being removed outlier based on spectral clustering according to claim 1, feature are existed
In:The kernel function is:
Wherein, t is the threshold constant of setting.
3. a kind of robust Model approximating method being removed outlier based on spectral clustering according to claim 1, feature are existed
In:The maximum iteration is 5 times.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810494460.2A CN108710761A (en) | 2018-05-22 | 2018-05-22 | A kind of robust Model approximating method removing outlier based on spectral clustering |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810494460.2A CN108710761A (en) | 2018-05-22 | 2018-05-22 | A kind of robust Model approximating method removing outlier based on spectral clustering |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108710761A true CN108710761A (en) | 2018-10-26 |
Family
ID=63868523
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810494460.2A Pending CN108710761A (en) | 2018-05-22 | 2018-05-22 | A kind of robust Model approximating method removing outlier based on spectral clustering |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108710761A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109961086A (en) * | 2019-01-28 | 2019-07-02 | 平安科技(深圳)有限公司 | Abnormal point ratio optimization method and device based on cluster and SSE |
CN110163298A (en) * | 2019-05-31 | 2019-08-23 | 闽江学院 | A kind of pattern fitting method of the sampling of fusant collection and model selection |
CN110163865A (en) * | 2019-05-28 | 2019-08-23 | 闽江学院 | A kind of method of sampling for unbalanced data in models fitting |
CN112132204A (en) * | 2020-09-18 | 2020-12-25 | 厦门大学 | Robust model fitting method based on preference probability weighted sampling |
US11113580B2 (en) | 2019-12-30 | 2021-09-07 | Industrial Technology Research Institute | Image classification system and method |
-
2018
- 2018-05-22 CN CN201810494460.2A patent/CN108710761A/en active Pending
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109961086A (en) * | 2019-01-28 | 2019-07-02 | 平安科技(深圳)有限公司 | Abnormal point ratio optimization method and device based on cluster and SSE |
CN109961086B (en) * | 2019-01-28 | 2024-05-31 | 平安科技(深圳)有限公司 | Clustering and SSE-based outlier proportion optimization method and device |
CN110163865A (en) * | 2019-05-28 | 2019-08-23 | 闽江学院 | A kind of method of sampling for unbalanced data in models fitting |
CN110163865B (en) * | 2019-05-28 | 2021-06-01 | 闽江学院 | Sampling method for unbalanced data in model fitting |
CN110163298A (en) * | 2019-05-31 | 2019-08-23 | 闽江学院 | A kind of pattern fitting method of the sampling of fusant collection and model selection |
US11113580B2 (en) | 2019-12-30 | 2021-09-07 | Industrial Technology Research Institute | Image classification system and method |
CN112132204A (en) * | 2020-09-18 | 2020-12-25 | 厦门大学 | Robust model fitting method based on preference probability weighted sampling |
CN112132204B (en) * | 2020-09-18 | 2022-05-24 | 厦门大学 | Robust model fitting method based on preference probability weighted sampling |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108710761A (en) | A kind of robust Model approximating method removing outlier based on spectral clustering | |
CN110097543B (en) | Hot-rolled strip steel surface defect detection method based on generation type countermeasure network | |
CN110837768B (en) | Online detection and identification method for rare animal protection | |
CN105844627B (en) | A kind of sea-surface target image background suppressing method based on convolutional neural networks | |
CN104063876B (en) | Interactive image segmentation method | |
CN112183153A (en) | Object behavior detection method and device based on video analysis | |
CN107045631A (en) | Facial feature points detection method, device and equipment | |
CN103415825A (en) | System and method for gesture recognition | |
CN109741268B (en) | Damaged image complement method for wall painting | |
Jung et al. | Rigid motion segmentation using randomized voting | |
CN105426929B (en) | Object shapes alignment device, object handles devices and methods therefor | |
CN108960329B (en) | Chemical process fault detection method containing missing data | |
CN107194344B (en) | Human behavior recognition method adaptive to bone center | |
JP2009140009A5 (en) | ||
CN104156979B (en) | Deviant Behavior online test method in a kind of video based on gauss hybrid models | |
CN103559689A (en) | Removal method for point cloud noise points | |
CN111429481B (en) | Target tracking method, device and terminal based on adaptive expression | |
CN112756324B (en) | Article cleaning method and device and terminal equipment | |
CN102708367A (en) | Image identification method based on target contour features | |
CN109409231B (en) | Multi-feature fusion sign language recognition method based on self-adaptive hidden Markov | |
CN105243352A (en) | 3D motion recognition method and device | |
CN106296747A (en) | Robust multi-model approximating method based on structure decision diagram | |
Tzagkarakis et al. | A statistical approach to texture image retrieval via alpha-stable modeling of wavelet decompositions | |
CN108256578B (en) | Gray level image identification method, device, equipment and readable storage medium | |
CN111723737B (en) | Target detection method based on multi-scale matching strategy deep feature learning |
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 |
Application publication date: 20181026 |
|
RJ01 | Rejection of invention patent application after publication |