CN110458241A - A kind of multi-angle of view classifier and its design method based on information enhancement - Google Patents
A kind of multi-angle of view classifier and its design method based on information enhancement Download PDFInfo
- Publication number
- CN110458241A CN110458241A CN201910759013.XA CN201910759013A CN110458241A CN 110458241 A CN110458241 A CN 110458241A CN 201910759013 A CN201910759013 A CN 201910759013A CN 110458241 A CN110458241 A CN 110458241A
- Authority
- CN
- China
- Prior art keywords
- angle
- sample
- information
- view
- data set
- 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
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/906—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of multi-angle of view classifier and its design method based on information enhancement, the multi-angle of view classifier include the multi-angle of view data collection module being sequentially connected, missing sample information repair module and effective sample information enhancement module.The present invention, for fields such as harbours, can effectively promote classification performance of the multi-angle of view data set in actual scene under relevant interface effect by two aspects such as missing sample information reparation, effective sample information enhancement.
Description
Technical field
The present invention relates to multi-angle of view learning art fields, and in particular to a kind of multi-angle of view classifier based on information enhancement and
Its design method.
Background technique
During making " smart city " comprehensively, people's data set to be treated often has many forms
Or source.This kind of data set is referred to as multi-angle of view data set, and a kind of form of expression or source are exactly a visual angle (such as web data
Text, image, the video of concentration), and different types of information included in any visual angle is then referred to as feature (such as text view
Textcolor, size text, text thickness in angle).Different from the single single-view data set of the form of expression or source, due to
This body structure of multi-angle of view data set is relative complex, so processing difficulty is higher, generally requires by being mentioned based on such data set
Multi-angle of view classifier out solves.In addition, being further divided into global characteristics drawn game for the feature of multi-angle of view data set
Portion's feature.The former is also referred to as coarseness feature, is mainly used for fuzzy matching and describes main feature information, such as profile, color
Etc. Global Informations；The latter is otherwise known as fine granularity feature, is mainly used for fine match and describes detailed information, as container is a certain
Somewhere information etc. on the special marking of position, the corresponding spectrum spectrogram of special biology.Since local feature is between sample
Fine difference is more sensitive, so current multi-angle of view classifier can more consideration is given to local features in design.
However, being limited to the objective factors such as sampling technique, human cost in the fields such as customs, harbour, traffic, can make
It presents and becomes privileged at the multi-angle of view data set of processing, specific manifestation are as follows: (1) visual angle or characteristic information missing: due to sampling technique
Limitation, people when acquiring multi-angle of view data set, can because human negligence or acquire equipment failure, cause part collected
Sample information on certain visual angles or feature is not complete, important so as to cause data set lack part to have classifier design
The visual angle of effect or characteristic information.For example, duration is carried out to an object with four cameras and shoots and records hypostome
To acquire data set, (in this example, the sample information of a camera forms a view to the characteristic informations such as color, size, profile
Angle).Because in certain time period provisional failure, which occurs, for a certain camera to work, then the object sample acquired in the period
Originally the information at a visual angle will be lost.If a certain camera, by electromagnetic interference, is directed to relevant viewing angles in certain time period,
The information acquired in the period will appear the case where Partial Feature is lost, and if profile is unintelligible, there is no obtain for size information
Record etc.；(2) have exemplar specific gravity too small: due to the limitation of human cost, for a large amount of true multi-angle of view data sets,
Only a fraction sample is marked in advance, and most of training sample for participating in classifier design does not obtain classification mark
Note.In general, the sample for obtaining label, which is referred to as, exemplar, they can be provided with knows conducive to the priori of classifier design
Know, the sample without label is referred to as unlabeled exemplars, and the priori knowledge that they are provided is less.Therefore, for true more views
For angular data collection, due to there is exemplar accounting lower, the effective sample information for causing them usually to have is less, and excessive
Unlabeled exemplars may interfere with classifier design again, be affected so as to cause classifier performance.
It is becomed privileged since in real scene, multi-angle of view data set is presented, to influence the automatic work of traditional classifier
Can, and be forced to need excessive manual intervention, to reduce the operating efficiency in real scene.
To handle these special multi-angle of view data sets, people (include association from traditional multi-angle of view classifier design thought
Same training, Multiple Kernel Learning, sub-space learning, the study of more matrixes, spacing uniformity etc.) it is converted to special multi-angle of view classifier and sets
Meter, and propose corresponding processing scheme.
(1) for missing visual angle or the multi-angle of view data set of feature: due to human negligence or the failure of acquisition equipment, people
When collecting multi-angle of view data set, it may appear that the missing of some visual angles or characteristic information.In order to solve such data set, phase
It closes scholar and proposes some solutions.For example, document [C.Xu, D.C.Tao, C.Xu, Multi-view learning with
incomplete views,IEEE Transactions on Image Processing,2015,24(12):5812-
5825.] it is directed to the multi-angle of view data set at imperfect visual angle, proposes and assumes that the information of matrix repairs algorithm based on low-rank.Document
[Q.Y.Yin,S.Wu,L.Wang,Unified subspace learning for incomplete and unlabeled
Multi-view data, Pattern Recognition, 2017,67:313-327.] research keep visual angle between and visual angle in
Characteristic similarity algorithm proposes the information based on unified sub-space learning and repairs algorithm.Document [L.Zhao, Z.K.Chen,
Y.Yang,Z.J.Wang,V.C.M.Leung,Incomplete multi-view clustering via deep
Semantic mapping, Neurocomputing, 2018,275:1053-1062.] propose based on deep semantic mapping and
The information of Affinity diagram repairs algorithm.The basic point of departure of these algorithms is all that the sample moment at each visual angle is obtained by Optimized model
Potential representation and corresponding coefficient matrix corresponding to battle array, using the product of the two to restore the information lost.
(2) for the multi-angle of view data set for having exemplar specific gravity too small: due to the limitation of cost of labor, for true field
For most of multi-angle of view data sets used in scape, ratio shared by the sample of label is just obtained before classifier training very
Small, this allows for related multi-angle of view classifier at the beginning of training, and the priori knowledge that can be obtained is extremely limited.Therefore for this kind of
There is multi-angle of view data set of the exemplar far fewer than unlabeled exemplars, related scholar proposes a series of algorithm, wherein commonly using
Be Universum learning series.Document [V.N.Vapnik, S.Kotz, Estimation of dependences based
On empirical data, 2006, Springer, New York, United States.] it points out, Universum study is calculated
Method by select other non-target class samples it is not intended that its class label or selection and merge two visual angles for having an exemplar,
Characteristic information, to obtain or generate a new unlabeled exemplars, i.e. Universum sample.The sample can be comprising certain
Priori knowledge.Currently, Universum study thoughts have derived to multiple fields.For example, document [X.H.Chen, H.J.Yin,
F.Jiang,L.P.Wang,Multi-view dimensionality reduction based on Universum
Learning, Neurocomputing, 2018,275:2279-2286.] propose that the typical association analysis based on Universum is calculated
Method, subspace unified expression is better achieved.Document [P.Songsiri, V.Cherkassky, B.Kijsirikul,
Universum selection for boosting the performance of multiclass support vector
machines based on one-versus-one strategy,Knowledge-Based Systems,2018,159:9-
19.] propose based on the Universum sample for being distributed symmetrically index (distributive and symmetric index, DSI)
This selection scheme is simultaneously used for support vector machines, to reduce the sample size for participating in the training of final classification device and while keep opposite
Higher classification performance.Document [B.Richhariya, D.Gupta, Facial expression recognition using
iterative universum twin support vector machine,Applied Soft Computing,2019,
76:53-67.] uncertainty based on sample, the generation of Universum sample and selection algorithm based on entropy are proposed, to make
The Universum sample that participation classifier training must be selected is representative.
These work have all preferably handled the multi-angle of view data set of some special shapes currently encountered.But pass through depth
Enter analysis, still find that it is existing following insufficient:
Missing sample information has to be repaired: although related scholar proposes a series of information repair mode, they
When restoration information, basic way is all to obtain potential expression corresponding to the sample matrix at each visual angle as much as possible
The product of the two is then considered as the information after repairing by form and corresponding coefficient matrix.But these methods are in restoration information
When, the amount of reparation is only taken into account, but ignores whether the information after repairing can bring higher recognition performance.In addition,
The corresponding same coefficient matrix in the multiple visual angles of part working hypothesis.This hypothesis can be convenient really and accelerate the reparation of information
Speed, but be unfavorable for preferably reflect different perspectives otherness.
Effective sample information has to be reinforced: having exemplar to set for providing effective sample information, being conducive to classifier
The priori knowledge of meter is critically important.But in actual life, this kind of sample is usually smaller in the accounting that each truthful data is concentrated.To the greatest extent
Pipe, some scholars propose a series of Universum learning algorithm, but present in these related algorithms centainly not
Foot.Such as do not account for original unlabeled exemplars effect (although they provide priori knowledge it is seldom, be not meant to
These samples are not worth), do not embody different perspectives or feature for effect of classifier design etc..
Summary of the invention
The object of the present invention is to provide a kind of multi-angle of view classifier and its design method based on information enhancement, passes through missing
Two aspects such as sample information reparation, effective sample information enhancement, under relevant interface effect, for fields such as harbours, to have
Effect promotes classification performance of the multi-angle of view data set in actual scene.
In order to achieve the above objectives, the present invention provides a kind of multi-angle of view classifier based on information enhancement comprising successively
Connected multi-angle of view data collection module, missing sample information repair module and effective sample information enhancement module；
The multi-angle of view data collection module is used to carry out zone location to the multi-angle of view data set of collection and label is handled,
And it is stored；
The missing sample information repair module is used for sample moment corresponding to each visual angle for multi-angle of view data set
Battle array calculates low-rank corresponding to sample matrix and assumes matrix and establish sub-classifier corresponding to the visual angle；Low-rank is assumed into square
Battle array is decomposed into the potential representation of sample matrix and coefficient matrix and updates sub-classifier；And then it obtains repairing letter for reflecting
Cease the amount expression formula of quantity and the matter expression formula of the classification performance for reflecting restoration information；Based on amount expression formula and matter expression formula
Building amount matter balance model, and then establish majorized function and the function is solved, obtain the potential expression shape at each visual angle
The optimum results of formula and coefficient matrix are multiplied to obtain the multi-angle of view data set after information is repaired by the two；
The effective sample information enhancement module is used to cluster and calculate using multi-angle of view for the multi-angle of view data set after repairing
Method is to obtain the weight at visual angle and feature；Calculate a phase having between exemplar and a unlabeled exemplars of any selection
Like degree；According to calculated similarity and a selection criteria, select have exemplar and nothing corresponding to suitable similarity
Exemplar, and suitable Universum sample is generated, so that enhancing is effectively used for the sample letter of multi-angle of view classifier training
Breath.
The above-mentioned multi-angle of view classifier based on information enhancement, wherein the multi-angle of view data collection module includes:
Image automatic positioning and label submodule, for carrying out zone location and mark to the multi-angle of view data set of collection
Reason；
Database purchase submodule is automatically positioned with described image and submodule is marked to be connected, and for storing, treated
Multi-angle of view data set.
The above-mentioned multi-angle of view classifier based on information enhancement, wherein the missing sample information repair module includes:
Meter operator module is assumed the expression formula of matrix for the low-rank according to corresponding to multi-angle of view data set, is calculated
To for reflecting the amount expression formula of restoration information quantity；
Matter computational submodule, the son point for the visual angle that the feature under each visual angle according to multi-angle of view data set is established
The matter expression formula for reflecting the classification performance of restoration information is calculated in class device；
Information repairs submodule, for establishing optimization to the amount matter balance model based on amount expression formula and the building of matter expression formula
Function simultaneously solves the function, obtains the potential representation at each visual angle and the optimum results of coefficient matrix, passes through two
Person is multiplied to obtain the multi-angle of view data set after information is repaired.
The above-mentioned multi-angle of view classifier based on information enhancement, wherein the effective sample information enhancement module includes:
Multi-angle of view clustering algorithm submodule, the multi-angle of view data set for being directed to after repairing, using multi-angle of view clustering algorithm
To obtain the weight at visual angle and feature；
Sample Similarity computational submodule is connected, for calculating any selection with the multi-angle of view clustering algorithm submodule
A similarity having between exemplar and a unlabeled exemplars；
Universum sample generates and selection submodule, is connected with the Sample Similarity computational submodule, is used for basis
Calculated similarity and a selection criteria select have exemplar and unlabeled exemplars corresponding to suitable similarity,
And suitable Universum sample is generated, so that enhancing is effectively used for the sample information of multi-angle of view classifier training.
The above-mentioned multi-angle of view classifier based on information enhancement, wherein the classifier is realized by Python
Model.
The design method of the present invention also provides a kind of multi-angle of view classifier based on information enhancement comprising following step
It is rapid:
Step 1: zone location being carried out to the multi-angle of view data set of collection and label is handled, and is stored；
Step 2: sample matrix corresponding to each visual angle for multi-angle of view data set calculates corresponding to sample matrix
Low-rank assumes matrix and establishes sub-classifier corresponding to the visual angle；Low-rank is assumed that matrix decomposition is the potential table of sample matrix
Show form and coefficient matrix and updates sub-classifier；And then it obtains the amount expression formula for reflecting restoration information quantity and is used for anti-
Reflect the matter expression formula of the classification performance of restoration information；Based on amount expression formula and matter expression formula building amount matter balance model, Jin Erjian
Vertical majorized function simultaneously solves the function, obtains the potential representation at each visual angle and the optimum results of coefficient matrix,
It is multiplied to obtain the multi-angle of view data set after information is repaired by the two；
Step 3: for the multi-angle of view data set after reparation, using multi-angle of view clustering algorithm to obtain the power at visual angle and feature
Weight；Calculate a similarity having between exemplar and a unlabeled exemplars of any selection；According to calculated similar
Degree and a selection criteria, select have exemplar and unlabeled exemplars corresponding to suitable similarity, and generate suitably
Universum sample, so that enhancing is effectively used for the sample information of multi-angle of view classifier training.
Compared with the existing technology, the invention has the following advantages:
The present invention is acted on by two aspects such as missing sample information reparation, effective sample information enhancement in relevant interface
Under, for fields such as harbours, it can effectively promote classification performance of the multi-angle of view data set in actual scene.
Detailed description of the invention
Fig. 1 is structural block diagram of the invention；
Fig. 2 is the working principle diagram of multi-angle of view data collection module in the present invention；
Fig. 3 is the working principle diagram that sample information repair module is lacked in the present invention；
Fig. 4 is the working principle diagram of effective sample information enhancement module in the present invention.
Specific embodiment
Below in conjunction with attached drawing, by specific embodiment, the invention will be further described, these embodiments are merely to illustrate
The present invention is not limiting the scope of the invention.
As shown in Figure 1, the invention discloses a kind of multi-angle of view classifier based on information enhancement, the classifier be by
Python realize model, it includes multi-angle of view data collection module 1, missing sample information repair module 2 and effectively
Sample information enhances module 3.In the present embodiment, the multi-angle of view data collection module 1 for including can be from UCI machine learning library
Multi-angle of view data set is collected in the scenes such as (http://archive.ics.uci.edu/ml/), true harbor service and by data
Collection sends missing sample information repair module 2 to, and then sends effective sample information enhancement module 3 to.
The multi-angle of view data collection module 1 is starting module, is done for collecting data set, and to the data set of collection
The processing such as positioning, label, and it is stored in database, so as to subsequent use.Specifically, first by be mounted on actual scene camera or
Other modes shooting obtains original image；Zone location and label are carried out to the picture of acquisition again, to lock spy to be processed
Reference ceases position；Finally, these characteristic informations are stored in database.
In the present embodiment, the multi-angle of view data collection module 1 includes: image automatic positioning and label submodule are used
Zone location and label are carried out in taking picture to hardware devices such as cameras, facilitates confirmation data set to be processed；Data
Library sub-module stored, the processing part for will obtain in original image stores, so as to subsequent use.
In the present embodiment, as shown in Fig. 2, firstly, using be mounted on actual scene camera or other modes shooting or
Obtain original image；Then, target positioning is carried out to the picture of acquisition, obtains target data information to be processed in whole picture
In relevant position；Then, target information is marked；Finally, will label result and relevant data set information deposit number
According to library, to obtain the data set of subsequent processing.
The missing sample information repair module 2 connects multi-angle of view data collection module 1, for the data after collection
Collection carries out the reparation of information.Specifically, firstly, being directed to sample moment corresponding to each visual angle of collected multi-angle of view data set
(for multi-angle of view data set, each of which visual angle has multiple features to battle array, and if web data is concentrated, text visual angle has text
The features such as this size, textcolor, text thickness.The characteristic information of different samples under the same perspective forms the sample at the visual angle
Matrix), it calculates its low-rank and assumes matrix and establish sub-classifier corresponding to the visual angle.Then, low-rank is assumed into matrix decomposition
For the potential representation of original sample matrix and corresponding coefficient matrix and the expression formula for updating sub-classifier.Then, it utilizes
Meter operator module 21 and matter computational submodule 22 establish an amount matter balance model to all visual angles.Finally, being repaired using information
After multiple submodule 23 optimizes the model, the potential representation at each visual angle and the optimum results of coefficient matrix are obtained, pass through two
Person obtains the data set after information is repaired after being multiplied.
In the present embodiment, the missing sample information repair module 2 includes: meter operator module 21, for according to receipts
Low-rank corresponding to the data set collected assumes the expression formula of matrix, and the amount for reflecting restoration information quantity is calculated and expresses
Formula；Matter computational submodule 22, the son point for the visual angle that the feature under each visual angle according to the data set being collected into is established
The matter expression formula for reflecting the classification performance of restoration information is calculated in class device；Information repairs submodule 23, for optimizing base
The amount matter balance model constructed by the amount expression formula and matter expression formula of restoration information, and optimized variable is obtained, thus realization pair
The reparation of missing information.
In the present embodiment, as shown in figure 3, firstly, for collect the multi-angle of view data set with v visual angle for, it is right
Each of which visual angle calculates low-rank corresponding to sample matrix and assumes matrix and initialize the corresponding sub-classifier in the visual angle, wherein
Ri indicates the sample matrix after information reparation.Then, low-rank is assumed that matrix decomposition is the latent of sample matrix using matrix decomposition
In representation and coefficient matrix and update sub-classifier.Since the present invention considers the source differences of different perspectives information,
So their coefficient matrix is also different.Then, meter is utilized to the potential representation and coefficient matrix at each visual angle
Operator module 21 obtains the amount expression formula for reflecting restoration information quantity, calculates son using matter to the sub-classifier at each visual angle
Module 22 obtains the matter expression formula for reflecting the classification performance of restoration information, and wherein Yi indicates classification matrix.Later, dosage matter
Balance model establishes majorized function.The function is solved finally, repairing submodule 23 using information, obtains each visual angle
The optimum results of potential representation and coefficient matrix, the information after being repaired by the two product, are repaired to obtain information
Multi-angle of view data set after multiple.
The connection missing sample information repair module 2 of effective sample information enhancement module 3, after for repairing
Data set further enhances sample information, and more suitably increases information by correlation criterion.Specifically, firstly, being directed to
Multi-angle of view data set after reparation, use multi-angle of view clustering algorithm submodule 31 (algorithm can with reference papers [Y.M.Xu,
C.D.Wang,J.H.Lai,Weighted multi-view clustering with feature selection,
Pattern Recognition, 2016,53:25-35.]) to obtain the weight of different perspectives, different characteristic.Then, one is selected
It is a to have exemplar and a unlabeled exemplars, according to visual angle, the weight of feature, counted by Sample Similarity computational submodule 32
Calculate the similarity between them.Finally, generating and selecting submodule 33 by Universum sample according to similarity and weight
To generate and select suitable Universum sample.These selected Universum samples and original there is label after repairing
Be formed together the data set that an effective sample information is enhanced with unlabeled exemplars, and participate in last classifier design and
Training.
In the present embodiment, the effective sample information enhancement module 3 includes: multi-angle of view clustering algorithm submodule 31, uses
Different perspectives, the weight of different characteristic are obtained in data set after repairing from information；Sample Similarity computational submodule 32 is used
In the weight according to calculated different perspectives, feature, calculating one of any selection has exemplar and one without label sample
Similarity between this；Universum sample generates and selection submodule 33, for being selected according to calculated similarity and one
Standard is selected, selects have exemplar and unlabeled exemplars corresponding to suitable similarity, and generates suitable Universum sample
This, so that enhancing is effectively used for the sample information of multi-angle of view classifier training.
In the present embodiment, as shown in figure 4, firstly, passing through multi-angle of view clustering algorithm for the data set after information reparation
Module 31 calculates and obtains different perspectives, the weight of feature, wherein wiIndicate weight, a at i-th of visual angleiIt indicates in i-th of visual angle
Weight vectors composed by the weight of each feature.Each numerical value of the weight vectors indicates the power of individual features under the visual angle
Weight.In this research contents, weight reflects the influence and effect of different perspectives, feature to classifier design and training.Then,
Based on calculating the visual angle obtained, feature weight, selecting one has exemplar xlWith a unlabeled exemplars xu, utilize sample
This similarity calculation submodule 32 obtains x using similarity calculation expression formulalAnd xuBetween similarity sl-u.Finally, according to
Similarity is generated and is selected submodule 33 by Universum sample, expression formula is generated first with Universum sample, to every
A pair has exemplar and unlabeled exemplars to calculate a Universum sample Ul-u, further according to Universum samples selection table
Up to formula, the operation such as sort to similarity, wherein the high Universum sample of similarity forms a selected set U for selections。
In the selection expression formula, sort indicates that sorting operation, select indicate selection operation, and therein selected
Universum sample is then considered as the sample with high-efficiency.Obtaining UsLater, after it being repaired with original information
Data set group becomes the data set that an effective sample information is enhanced, and finally participates in the design and training of classifier.
The multi-angle of view classifier design method based on information enhancement that the invention also discloses a kind of, it includes following steps:
Step 1: using hardware devices such as the cameras being installed in different actual scenes, photographic subjects picture is simultaneously determined
Position, label processing, so that storage is into database so as to subsequent use；
Step 2: matrix and visual angle sub-classifier, building are assumed based on low-rank corresponding to the multi-angle of view data set being collected into
Amount matter balance model is simultaneously subject to Optimization Solution, obtains the data set after information is repaired；
Step 3: the data set after being repaired based on information acquisition seeks the weight at visual angle, feature, and according to the phase between sample
Like degree, to generate and select suitable Universum sample to participate in the design of classifier, to enhance effective sample information；
By in the application system of the classifier insertion related fields of design, multi-angle of view number is handled to be lifted in practical field
According to the performance of collection.
In conclusion in terms of the present invention is by missing sample information reparation, effective sample information enhancement etc. two, in correlation
Under interface effect, for fields such as harbours, classification performance of the multi-angle of view data set in actual scene can be effectively promoted.
It is discussed in detail although the contents of the present invention have passed through above preferred embodiment, but it should be appreciated that above-mentioned
Description is not considered as limitation of the present invention.After those skilled in the art have read above content, for of the invention
A variety of modifications and substitutions all will be apparent.Therefore, protection scope of the present invention should be limited to the appended claims.
Claims (6)
1. a kind of multi-angle of view classifier based on information enhancement, which is characterized in that including the multi-angle of view data collection being sequentially connected
Module, missing sample information repair module and effective sample information enhancement module；
The multi-angle of view data collection module is used to carry out zone location to the multi-angle of view data set of collection and label is handled, and goes forward side by side
Row storage；
The missing sample information repair module is used for sample matrix corresponding to each visual angle for multi-angle of view data set, meter
Low-rank corresponding to sample matrix is calculated to assume matrix and establish sub-classifier corresponding to the visual angle；Low-rank is assumed into matrix decomposition
For sample matrix potential representation and coefficient matrix and update sub-classifier；And then it obtains for reflecting restoration information quantity
Amount expression formula and the classification performance for reflecting restoration information matter expression formula；Based on amount expression formula and matter expression formula building amount
Matter balance model, and then establish majorized function and the function is solved, obtain the potential representation at each visual angle and be
The optimum results of matrix number are multiplied to obtain the multi-angle of view data set after information is repaired by the two；
The effective sample information enhancement module be used for for repair after multi-angle of view data set, use multi-angle of view clustering algorithm with
Obtain the weight at visual angle and feature；Calculate one of any selection have it is similar between exemplar and a unlabeled exemplars
Degree；According to calculated similarity and a selection criteria, select have exemplar corresponding to suitable similarity and without mark
Signed-off sample sheet, and suitable Universum sample is generated, so that enhancing is effectively used for the sample letter of multi-angle of view classifier training
Breath.
2. the multi-angle of view classifier based on information enhancement as described in claim 1, which is characterized in that the multi-angle of view data are received
Collecting module includes:
Image automatic positioning and label submodule, for carrying out zone location and label processing to the multi-angle of view data set of collection；
Database purchase submodule is connected with described image automatic positioning and label submodule, and for storing, treated regards more
Angular data collection.
3. the multi-angle of view classifier based on information enhancement as described in claim 1, which is characterized in that the missing sample information
Repair module includes:
Meter operator module assumes the expression formula of matrix for the low-rank according to corresponding to multi-angle of view data set, use is calculated
In the amount expression formula of reflection restoration information quantity；
Matter computational submodule, the subclassification for the visual angle that the feature under each visual angle according to multi-angle of view data set is established
The matter expression formula for reflecting the classification performance of restoration information is calculated in device；
Information repairs submodule, for establishing majorized function to the amount matter balance model based on amount expression formula and the building of matter expression formula
And the function is solved, the potential representation at each visual angle and the optimum results of coefficient matrix are obtained, the two phase is passed through
Multi-angle of view data set after the multiplied reparation to information.
4. the multi-angle of view classifier based on information enhancement as described in claim 1, which is characterized in that the effective sample information
Enhancing module includes:
Multi-angle of view clustering algorithm submodule, for using multi-angle of view clustering algorithm to obtain for the multi-angle of view data set after repairing
Obtain the weight at visual angle and feature；
Sample Similarity computational submodule is connected, for calculating the one of any selection with the multi-angle of view clustering algorithm submodule
A similarity having between exemplar and a unlabeled exemplars；
Universum sample generates and selection submodule, is connected with the Sample Similarity computational submodule, for according to calculating
Similarity and a selection criteria out, select have exemplar and unlabeled exemplars corresponding to suitable similarity, and raw
At suitable Universum sample, so that enhancing is effectively used for the sample information of multi-angle of view classifier training.
5. the multi-angle of view classifier based on information enhancement as described in claim 1, which is characterized in that the classifier be by
The model that Python is realized.
6. a kind of design method of the multi-angle of view classifier based on information enhancement, which comprises the following steps:
Step 1: zone location being carried out to the multi-angle of view data set of collection and label is handled, and is stored；
Step 2: sample matrix corresponding to each visual angle for multi-angle of view data set calculates low-rank corresponding to sample matrix
Assuming that matrix and establishing sub-classifier corresponding to the visual angle；Low-rank is assumed that matrix decomposition is the potential expression shape of sample matrix
Formula and coefficient matrix simultaneously update sub-classifier；And then it obtains the amount expression formula for reflecting restoration information quantity and is repaired for reflecting
The matter expression formula of the classification performance of complex information；Based on amount expression formula and matter expression formula building amount matter balance model, and then establish excellent
Change function and the function is solved, obtains the potential representation at each visual angle and the optimum results of coefficient matrix, pass through
The two is multiplied to obtain the multi-angle of view data set after information is repaired；
Step 3: for the multi-angle of view data set after reparation, using multi-angle of view clustering algorithm to obtain the weight at visual angle and feature；
Calculate a similarity having between exemplar and a unlabeled exemplars of any selection；According to calculated similarity and
One selection criteria selects have exemplar and unlabeled exemplars corresponding to suitable similarity, and generates suitably
Universum sample, so that enhancing is effectively used for the sample information of multi-angle of view classifier training.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910759013.XA CN110458241A (en) | 2019-08-16 | 2019-08-16 | A kind of multi-angle of view classifier and its design method based on information enhancement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910759013.XA CN110458241A (en) | 2019-08-16 | 2019-08-16 | A kind of multi-angle of view classifier and its design method based on information enhancement |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110458241A true CN110458241A (en) | 2019-11-15 |
Family
ID=68487243
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910759013.XA Pending CN110458241A (en) | 2019-08-16 | 2019-08-16 | A kind of multi-angle of view classifier and its design method based on information enhancement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110458241A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111047052A (en) * | 2019-12-24 | 2020-04-21 | 上海海事大学 | Semi-supervised multi-view data set online learning model and design method thereof |
| CN111814554A (en) * | 2020-06-09 | 2020-10-23 | 同济大学 | Object type recognition model construction method based on granularity and associated information and application |
| WO2021244528A1 (en) * | 2020-06-05 | 2021-12-09 | 上海海事大学 | Information enhancement method and information enhancement system |
-
2019
- 2019-08-16 CN CN201910759013.XA patent/CN110458241A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111047052A (en) * | 2019-12-24 | 2020-04-21 | 上海海事大学 | Semi-supervised multi-view data set online learning model and design method thereof |
| WO2021244528A1 (en) * | 2020-06-05 | 2021-12-09 | 上海海事大学 | Information enhancement method and information enhancement system |
| CN111814554A (en) * | 2020-06-09 | 2020-10-23 | 同济大学 | Object type recognition model construction method based on granularity and associated information and application |
| CN111814554B (en) * | 2020-06-09 | 2022-06-21 | 同济大学 | Object type recognition model construction method based on granularity and associated information and application |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106157307B (en) | A kind of monocular image depth estimation method based on multiple dimensioned CNN and continuous CRF | |
| CN109522853B (en) | Face datection and searching method towards monitor video | |
| Liu et al. | A smart unstaffed retail shop based on artificial intelligence and IoT | |
| CN110458241A (en) | A kind of multi-angle of view classifier and its design method based on information enhancement | |
| CN102932605B (en) | Method for selecting camera combination in visual perception network | |
| CN110490238A (en) | A kind of image processing method, device and storage medium | |
| CN109214263A (en) | A kind of face identification method based on feature multiplexing | |
| CN109145717A (en) | A kind of face identification method of on-line study | |
| CN107798313A (en) | A kind of human posture recognition method, device, terminal and storage medium | |
| CN107038400A (en) | Face identification device and method and utilize its target person tracks of device and method | |
| CN110163567A (en) | Classroom roll calling system based on multitask concatenated convolutional neural network | |
| CN111539351A (en) | Multi-task cascaded face frame selection comparison method | |
| Zhang et al. | Masked face recognition with mask transfer and self-attention under the COVID-19 pandemic | |
| CN110580510A (en) | clustering result evaluation method and system | |
| CN109815823A (en) | Data processing method and Related product | |
| JP2000322577A (en) | Device and method for collating image and recording medium with its control program recorded thereon | |
| CN109920050A (en) | A kind of single-view three-dimensional flame method for reconstructing based on deep learning and thin plate spline | |
| CN111126155A (en) | Pedestrian re-identification method for generating confrontation network based on semantic constraint | |
| Geng et al. | Shelf Product Detection Based on Deep Neural Network | |
| CN109064578A (en) | A kind of attendance system and method based on cloud service | |
| CN115588220A (en) | Two-stage multi-scale self-adaptive low-resolution face recognition method and application | |
| CN108596068A (en) | A kind of method and apparatus of action recognition | |
| Chen et al. | Poker Watcher: Playing Card Detection Based on EfficientDet and Sandglass Block | |
| WO2021129491A1 (en) | Pedestrian search method, server, and storage medium | |
| CN111047052A (en) | Semi-supervised multi-view data set online learning model and design method thereof |
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: 20191115 |
|
| RJ01 | Rejection of invention patent application after publication |