CN107203786A - Image-recognizing method and device based on sparse border Fisher algorithms - Google Patents
Image-recognizing method and device based on sparse border Fisher algorithms Download PDFInfo
- Publication number
- CN107203786A CN107203786A CN201710425040.4A CN201710425040A CN107203786A CN 107203786 A CN107203786 A CN 107203786A CN 201710425040 A CN201710425040 A CN 201710425040A CN 107203786 A CN107203786 A CN 107203786A
- Authority
- CN
- China
- Prior art keywords
- training dataset
- projection matrix
- matrix
- sparse
- test data
- 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
- 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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2132—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on discrimination criteria, e.g. discriminant analysis
- G06F18/21322—Rendering the within-class scatter matrix non-singular
-
- 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
-
- 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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2132—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on discrimination criteria, e.g. discriminant analysis
- G06F18/21322—Rendering the within-class scatter matrix non-singular
- G06F18/21324—Rendering the within-class scatter matrix non-singular involving projections, e.g. Fisherface techniques
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image-recognizing method and device based on sparse border Fisher algorithms, this method includes：The first training dataset is obtained, projection matrix is obtained according to the first training dataset；LS-SVM sparseness is carried out to projection matrix, sparse projection matrix is obtained；First training dataset by sparse projection matrix project to obtain the second training dataset；The first test data of images to be recognized is received, the first test data by sparse projection matrix project to obtain the second test data；The second test data is identified on the second training dataset using sorting algorithm.Therefore, by to training dataset project obtaining projection matrix, can maximum retention data geometry, projection matrix is subjected to LS-SVM sparseness again, more preferable generalization ability and discriminating power can be obtained, the problem of solving the study disequilibrium of the differentiation to all data, while improving discrimination.
Description
Technical field
The present invention relates to field of image recognition, more specifically to a kind of figure based on sparse border Fisher algorithms
As recognition methods and device.
Background technology
With the development of artificial intelligence technology, the form of man-machine interaction constantly expands abundant, builds effective man-machine interaction
As development trend.Face be human emotion expression and most important, the most direct mode exchanged, therefore face recognition technology by
Gradually turn into a focus of area of pattern recognition in man-machine interaction.Recognition methods based on face characteristic also gradually grows up,
But the mature and reliable not enough of the extracting method to face characteristic used.
Fisher analysis (MFA) algorithms in border are a kind of feature extracting methods, and it is a kind of figure of combination class boundary information
Embedded mobile GIS, basic thought is one intrinsic figure of construction, makes sample point in class compacter, while one punishment figure of construction, makes
Boundary sample point between foreign peoples is more separated, and takes full advantage of the local analog information in border discriminant information and class between class, will
The geometry of data retains well, but is due to the global discriminant information that local differentiation structure have ignored data set, can
The differentiation of all data can be caused to learn disequilibrium.
Therefore, how well retention data geometry while ensure data set global discriminant information be this area
The problem of technical staff needs to solve.
The content of the invention
It is an object of the invention to provide a kind of feature extracting method analyzed based on border Fisher, to retain well
Data geometry ensures the global discriminant information of data set simultaneously, improves discrimination.
To achieve the above object, the embodiments of the invention provide following technical scheme：
A kind of image-recognizing method based on sparse border Fisher algorithms, including：
The first training dataset is obtained, projection matrix is obtained according to first training dataset；
LS-SVM sparseness is carried out to the projection matrix, sparse projection matrix is obtained；
First training dataset is projected by the sparse projection matrix, the second training dataset is obtained；
The first test data of images to be recognized is received, first test data is entered by the sparse projection matrix
Row projection obtains the second test data；
Second test data is identified on second training dataset using sorting algorithm.
Preferably, it is described that second test data is known on second training dataset using sorting algorithm
Not, including：
Second test data is identified on second training dataset using arest neighbors sorting algorithm.
Preferably, the first training dataset of the acquisition, projection matrix is obtained according to first training dataset, is wrapped
Include：
Obtain the first training datasetWherein xi∈RD, yi∈ { 1,2 ..., c }, N are the number of sample, D
It is the dimension of sample, yiIt is xiClass label, c be classification sum；
Object function to finding projection matrix
Solve, obtain the projection matrix；Wherein, projection matrix is P and P ∈ RD×m, m is the subspace size of setting；X
∈RD×N, it is the sample matrix that the training data of the first training data concentration is made；Lb∈RN×N, it is Laplce's square between class
Battle array；Lw∈RN×N, it is Laplacian Matrix in class；A is to matrix X (Lw+Lb)XTBy its nonzero eigenvalue correspondence after feature decomposition
Characteristic vector composition matrix.
Preferably, LS-SVM sparseness is carried out to the projection matrix, obtains sparse projection matrix, including：
Solved using linear Bregman alternative mannersTo the projection matrix P rarefactions, obtain described
Sparse projection matrix V, wherein | | V | |1It is 1 norm.
Preferably, it is described to be projected first training dataset by the sparse projection matrix, obtain second
Training dataset, including：
First training datasetAccording to the sparse projection matrix V project obtaining the second training
Data setWherein zi=VTxi。
A kind of pattern recognition device based on sparse border Fisher algorithms, including：
Projection matrix acquisition module, for obtaining the first training dataset, is thrown according to first training dataset
Shadow matrix；
LS-SVM sparseness module, for carrying out LS-SVM sparseness to the projection matrix, obtains sparse projection matrix；
Training dataset projection module, for first training dataset to be thrown by the sparse projection matrix
Shadow, obtains the second training dataset；
Test data projection module, the first test data for receiving images to be recognized, to first test data
By the sparse projection matrix project obtaining the second test data；
Identification module, for being carried out using sorting algorithm to second test data on second training dataset
Identification.
Preferably, the identification module is specifically for utilizing arest neighbors sorting algorithm to second test data described
It is identified on second training dataset.
Preferably, the projection matrix acquisition module, including：
First training dataset acquiring unit, for obtaining the first training datasetWherein xi∈RD, yi∈
{ 1,2 ..., c }, N is the number of sample, and D is the dimension of sample, yiIt is xiClass label, c be classification sum；
Projection matrix acquiring unit, for the object function to finding projection matrixAsk
Solution, obtains the projection matrix；Wherein, projection matrix is P and P ∈ RD×m, m is the subspace size of setting；X∈RD×N, it is institute
State the sample matrix that the training data of the first training data concentration is made；Lb∈RN×N, it is Laplacian Matrix between class；Lw∈RN ×N, it is Laplacian Matrix in class；A is to matrix X (Lw+Lb)XTAfter feature decomposition from the corresponding feature of its nonzero eigenvalue to
Measure the matrix of composition.
Preferably, the LS-SVM sparseness module, specifically for being solved using linear Bregman alternative mannersTo the projection matrix P rarefactions, the sparse projection matrix V is obtained, wherein | | V | |1It is 1 norm.
Preferably, the training dataset projection module, specifically for first training datasetAccording to
The sparse projection matrix V project obtaining the second training datasetWherein zi=VTxi。
By above scheme, a kind of image based on sparse border Fisher algorithms provided in an embodiment of the present invention is known
Other method, including：The first training dataset is obtained, projection matrix is obtained according to first training dataset；To the projection
Matrix carries out LS-SVM sparseness, obtains sparse projection matrix；First training dataset is passed through into the sparse projection matrix
Projected, obtain the second training dataset；The first test data of images to be recognized is received, it is logical to first test data
The sparse projection matrix is crossed project obtaining the second test data；Using sorting algorithm to second test data in institute
State and be identified on the second training dataset.
Therefore, projection matrix is obtained by being projected to training dataset progress, that is, MFA is carried out to training dataset
Space learning, can make full use of the local analog information in border discriminant information and class between class, can retention data well
Geometry；Projection matrix is subjected to LS-SVM sparseness again, more preferable generalization ability and discriminating power can be obtained, it is ensured that data
The global discriminant information of collection, then by the first test data project the second obtained test data, then pass through sorting algorithm
It is identified, therefore had both been taken full advantage of between class for the identification of test data on the second training dataset after sparse projection
Local analog information in discriminant information and class, farthest remains the collecting structure of data, while also ensure that data
The problem of global discriminant information of collection solves the study disequilibrium of the differentiation to all data, improves discrimination；The present invention
A kind of pattern recognition device based on sparse border Fisher algorithms is also disclosed, above-mentioned technique effect can be equally realized.
Brief description of the drawings
In order to illustrate more clearly about the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing
There is the accompanying drawing used required in technology description to be briefly described, it should be apparent that, drawings in the following description are only this
Some embodiments of invention, for those of ordinary skill in the art, on the premise of not paying creative work, can be with
Other accompanying drawings are obtained according to these accompanying drawings.
Fig. 1 is a kind of image-recognizing method flow chart disclosed in the embodiment of the present invention；
Fig. 2 is a kind of pattern recognition device structural representation disclosed in the embodiment of the present invention；
Fig. 3 is a specific projection matrix modular structure schematic diagram disclosed in the embodiment of the present invention.
Embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete
Site preparation is described, it is clear that described embodiment is only a part of embodiment of the invention, rather than whole embodiments.It is based on
Embodiment in the present invention, it is every other that those of ordinary skill in the art are obtained under the premise of creative work is not made
Embodiment, belongs to the scope of protection of the invention.
The embodiment of the invention discloses a kind of feature extracting method analyzed based on border Fisher, to ensure data set
Global discriminant information, improves discrimination.
Referring to Fig. 1, a kind of image-recognizing method based on border Fisher algorithms provided in an embodiment of the present invention, including：
S101, obtains the first training dataset, projection matrix is obtained according to first training dataset.
Specifically, MFA (border Fisher analyses) sub-space learning is carried out to training image first, i.e., in training image
The first training dataset is obtained, and projection matrix is obtained according to training dataset.Wherein, projection matrix can be according to training number
Obtain finding the object function of projection matrix according to collection, projection matrix is obtained by being solved to object function.
It should be noted that carrying out editing discriminant information MFA sub-space learnings can make full use of class to training figure
With the local analog information in class, obtained projection matrix can farthest retention data geometry, it is but local
Differentiate the global discriminant information that have ignored data set.
S102, carries out LS-SVM sparseness to the projection matrix, obtains sparse projection matrix；
Specifically, LS-SVM sparseness is carried out to the projection matrix, obtains sparse projection matrix, original variable can be controlled
And feature weight, more preferable generalization ability and discriminating power can be obtained, MFA sub-space learnings is solved and ignores the complete of data set
The problem of office's discriminant information.
S103, first training dataset is projected by the sparse projection matrix, obtains the second training number
According to collection；
In this programme, the sparse projection matrix obtained by S102 is projected to the first training dataset, is thrown
The second data set of movie queen；It should be noted that the training dataset of needs in the projected is i.e. when classifying to test data
Classified on second training dataset.
S104, receives the first test data of images to be recognized, the sparse projection is passed through to first test data
Matrix project obtaining the second test data；
Specifically, test data, i.e. the first test data are obtained in images to be recognized first, recycles what S102 was obtained
Factor projection matrix is projected to the first test data, the test data after being projected, i.e. the second test data.
S105, second test data is identified on second training dataset using sorting algorithm.
Specifically, it is the training dataset of the second test data in the projected i.e. the second training by the test data after projection
Classified on data set, classification is realized by sorting algorithm, that is, test image data are identified, draw identification knot
Really.
Therefore, projection matrix is obtained by being projected to training dataset progress, that is, MFA is carried out to training dataset
Space learning, can make full use of the local analog information in border discriminant information and class between class, therefore can retain well
Data geometry；Projection matrix is subjected to LS-SVM sparseness again, more preferable generalization ability and discriminating power can be obtained, it is ensured that
The global discriminant information of data set, then by the first test data project the second obtained test data, then pass through classification
Algorithm is identified on the second training dataset after sparse projection, therefore, and the identification for test data had both made full use of
Local analog information between class in discriminant information and class, farthest remains the geometry of data, while also ensureing
The problem of global discriminant information of data set solves differentiation to all data study disequilibrium, improves discrimination.
The embodiment of the invention discloses a kind of specific image-recognizing method based on sparse border Fisher algorithms, relatively
In a upper embodiment, the present embodiment has done further instruction and optimization to technical scheme.Including：
S201, obtains the first training datasetWherein xi∈RD, yi∈ { 1,2 ..., c }, N is sample
Number, D is the dimension of sample, yiIt is xiClass label, c be classification sum；
Specifically, training dataset, i.e. the first training dataset are obtained in training image
S202, the object function to finding projection matrixSolve, obtain the projection square
Battle array；Wherein, projection matrix is P and P ∈ RD×m, m is the subspace size of setting；X∈RD×N, it is first training dataset
In training data composition sample matrix；Lb∈RN×N, it is Laplacian Matrix between class；Lw∈RN×N, it is Laplce in class
Matrix；A is to matrix X (Lw+Lb)XTThe matrix being made up of after feature decomposition the corresponding characteristic vector of its nonzero eigenvalue.
Specifically, the first training dataset obtained according to S201Obtain the instruction of the first training data concentration
Practice the sample matrix X of data composition, X is brought into the object function for finding projection matrix
It is middle to be solved, obtain the projection matrix P in object function.
S203, is solved using linear Bregman alternative mannersTo the projection matrix P rarefactions, obtain
The sparse projection matrix V, wherein | | V | |1It is 1 norm.
Specifically, by projection matrix rarefaction, solved using linear Bregman alternative mannersSo as to
Sparse projection matrix V after to projection matrix rarefaction.
LS-SVM sparseness is carried out by linear Bregman iteration, the weight of original variable and feature can be controlled, obtained more
Generalization ability and discriminating power well.
S204, first training datasetAccording to the sparse projection matrix V project obtaining second
Training datasetWherein zi=VTxi。
Specifically, the sparse projection matrix V obtained by S203 is to the first training datasetProjected,
The second data set after being projectedWherein ziBy xiCalculated by sparse projection matrix V come i.e. zi=VTxi。
S205, receives the first test data of images to be recognized, the sparse projection is passed through to first test data
Matrix project obtaining the second test data；
Specifically, the first test data x, wherein x ∈ R of figure to be identified are receivedD, LS-SVM sparseness is carried out to x, that is,
Say and projected x by sparse projection matrix V, obtain the second test data z, wherein z ∈ Rm。
S206, second test data is identified on second training dataset using sorting algorithm.
Specifically, using sorting algorithm to the second test data z in the second training datasetIt is upper to be divided
Class, that is, test image data are identified, draw recognition result.
Therefore, the embodiment of the present invention passes through to training datasetProgress, which is projected, obtains projection matrix P, that is,
MFA sub-space learnings are carried out to training dataset, discriminant information letter similar with the part in class in border between class can be made full use of
Breath, the geometry of maximum retention data；Projection matrix P is carried out at rarefaction by Bregman alternative manners again
Reason obtains sparse projection matrix V, can obtain more preferable generalization ability and discriminating power；Then the first test data x is carried out
Project the second obtained test data z, then the second training dataset by sorting algorithm after sparse projectionOn
It is identified, therefore local analog information between class in discriminant information and class had both been taken full advantage of for the identification of test data,
The problem of also solving the study disequilibrium of the differentiation to all data simultaneously, improves discrimination.
The embodiment of the invention discloses a kind of image-recognizing method for being specifically based on sparse border Fisher algorithms,
Relative to upper embodiment the present embodiment in a upper embodiment using sorting algorithm to second test data described the
It is identified on two training datasets and has done specific restriction, other steps is roughly the same with a upper embodiment, and detailed content can
So that referring to the corresponding part of a upper embodiment, here is omitted.Specifically：
Second test data is identified on second training dataset using arest neighbors sorting algorithm.
A kind of pattern recognition device based on sparse border Fisher algorithms provided in an embodiment of the present invention is carried out below
Introduce, can be cross-referenced with a kind of above-described image-recognizing method based on sparse border Fisher algorithms.
Referring to Fig. 2, the embodiment of the present invention provides a kind of pattern recognition device based on sparse border Fisher algorithms, including
Projection matrix acquisition module 301, LS-SVM sparseness module 302, training dataset projection module 303, test data projection module
304, identification module 305.
Projection matrix acquisition module 301, for obtaining the first training dataset, is obtained according to first training dataset
Projection matrix.
Specifically, projection matrix acquisition module 301 carries out MFA sub-space learnings to training image first, in training image
The first training dataset of middle acquisition, and projection matrix is obtained according to training dataset.Wherein it is possible to be obtained according to training dataset
Projection matrix is obtained to the object function for finding projection matrix, then by being solved to object function.
It should be noted that carrying out editing discriminant information MFA sub-space learnings can make full use of class to training figure
With the local analog information in class, obtained projection matrix can farthest retention data geometry, it is but local
Differentiate the global discriminant information that have ignored data set.
LS-SVM sparseness module 302, for carrying out LS-SVM sparseness to the projection matrix, obtains sparse projection matrix.
Specifically, 302 pairs of the LS-SVM sparseness module projection matrix carries out LS-SVM sparseness, obtains sparse projection square
Battle array, can control original variable and feature weight, can obtain more preferable generalization ability and discriminating power, solve MFA empty
Between study the problem of ignore the global discriminant information of data set.
Training dataset projection module 303, for first training dataset to be entered by the sparse projection matrix
Row projection, obtains the second training dataset.
Specifically, the sparse projection matrix obtained by LS-SVM sparseness module 302 is thrown the first training dataset
Shadow, the second training dataset after being projected.
Test data projection module 304, the first test data for receiving images to be recognized, to the described first test number
The second test data is obtained according to being projected by sparse projection matrix progress；
Specifically, test data projection module 304 obtains test data, i.e., the first test number in images to be recognized first
According to the sparse projection matrix for recycling LS-SVM sparseness module 302 to obtain is projected to the first test data, is obtained after projection
Test data, i.e. the second test data.
Identification module 305, for utilizing sorting algorithm to second test data in second training dataset
It is identified.
Specifically, the second test data test data projection module 304 obtained is in training dataset projection module 303
The second training dataset on classified, classification by sorting algorithm realization, that is, test image data are identified,
Draw recognition result.
Therefore, projection matrix acquisition module 301 obtains projection matrix by training dataset project, that is, right
Training dataset carries out MFA sub-space learnings, can make full use of the local analog information in border discriminant information and class between class,
The geometry of maximum retention data；Projection matrix is carried out LS-SVM sparseness by LS-SVM sparseness module 302 again, can be with
More preferable generalization ability and discriminating power are obtained, then test data projection module 304 carries out the first test data to project
The second test data arrived, identification module 305 is identified by sorting algorithm on the second training dataset again.Therefore, lead to
Cross and training dataset project to obtain projection matrix, that is, MFA sub-space learnings, Ke Yichong are carried out to training dataset
, can retention data geometry well point using the local analog information in border discriminant information and class between class；It will throw again
Shadow matrix carries out LS-SVM sparseness, can obtain more preferable generalization ability and discriminating power, it is ensured that the global of data set differentiates letter
Breath, then by the first test data project the second obtained test data, then by sorting algorithm after sparse projection
It is identified, therefore had both been taken full advantage of between class in discriminant information and class for the identification of test data on second training dataset
Local analog information, the collecting structure of data is farthest remained, while also ensure that the global of data set differentiates letter
The problem of breath solves the study disequilibrium of the differentiation to all data, improves discrimination.
The embodiment of the invention discloses a kind of pattern recognition device for being specifically based on sparse border Fisher algorithms,
Relative to a upper embodiment, the present embodiment has done further instruction and optimization to technical scheme.Specifically：
Projection matrix acquisition module 401, referring to Fig. 3, projection matrix acquisition module 401 is specifically included：
First training dataset acquiring unit 401a, for obtaining the first training datasetWherein xi∈RD,
yi∈ { 1,2 ..., c }, N are the numbers of sample, and D is the dimension of sample, yiIt is xiClass label, c be classification sum；
Specifically, the first training dataset acquiring unit obtains training dataset, i.e., the first training number in training image
According to collection
Projection matrix acquiring unit 401b, for the object function to finding projection matrix
Solve, obtain the projection matrix；Wherein, projection matrix is P and P ∈ RD×m, m is the subspace size of setting；X∈RD×N, it is
The sample matrix that the training data that first training data is concentrated is made；Lb∈RN×N, it is Laplacian Matrix between class；Lw∈RN ×N, it is Laplacian Matrix in class；A is to matrix X (Lw+Lb)XTAfter feature decomposition from the corresponding feature of its nonzero eigenvalue to
Measure the matrix of composition.
Specifically, the first training dataset obtained according to the first training dataset acquiring unit 401a
The sample matrix X for the training data composition concentrated to the first training data, X is brought into the object function for finding projection matrixIt is middle to be solved, obtain the projection matrix P in object function.
LS-SVM sparseness module 402, for being solved using linear Bregman alternative mannersThrow described
Shadow matrix P rarefactions, obtain the sparse projection matrix V, wherein | | V | |1It is 1 norm.
Specifically, by projection matrix rarefaction, solved using linear Bregman alternative mannersTo obtain
Sparse projection matrix V after projection matrix rarefaction.
LS-SVM sparseness is carried out by linear Bregman iteration, the weight of original variable and feature can be controlled, obtained more
Generalization ability and discriminating power well.
Training dataset projection module 403, for first training datasetAccording to the sparse projection
Matrix V project obtaining the second training datasetWherein zi=VTxi。
Specifically, the sparse projection matrix V obtained by LS-SVM sparseness module 402 is to the first training datasetProjected, the second data set after being projectedWherein ziBy xiPass through sparse projection matrix V
Calculate and, i.e. zi=VTxi。
Test data projection module 404, the first test data for receiving images to be recognized, to the described first test number
The second test data is obtained according to being projected by sparse projection matrix progress；
Specifically, the first test data x, wherein x ∈ R of figure to be identified are receivedD, LS-SVM sparseness is carried out to x, that is,
Say and projected x by sparse projection matrix V, obtain the second test data z, wherein z ∈ Rm。
Identification module 405, for utilizing sorting algorithm to second test data in second training dataset
It is identified.
Specifically, using sorting algorithm to the second test data z in the second training datasetIt is upper to be classified,
Test data is identified, recognition result is drawn.
Therefore, the first training dataset acquiring unit of embodiment of the present invention 401a passes through to training datasetEnter
Row projection obtains projection matrix P, that is, carries out MFA sub-space learnings to training dataset, can make full use of border between class
Local analog information in discriminant information and class, the geometry of maximum retention data；LS-SVM sparseness module 402 is again
Projection matrix P is subjected to LS-SVM sparseness by Bregman alternative manners and obtains sparse projection matrix V, can be obtained preferably
Generalization ability and discriminating power；Then the first test data x project the second obtained survey by test data projection module 404
Data z is tried, identification module 405 is again by sorting algorithm in the second training datasetOn be identified.Therefore for
The identification of test data had both taken full advantage of the local analog information in discriminant information and class between class, farthest remained number
According to collecting structure, while also ensure that data set global discriminant information solve the differentiation to all data study lose flat
The problem of weighing apparatus, improve discrimination.
The embodiment of the invention discloses a kind of pattern recognition device based on sparse border Fisher algorithms, relative to upper one
Embodiment the present embodiment has done specifically defined to identification module in a upper embodiment, other modules and a upper embodiment substantially phase
Together, detailed content may refer to the corresponding part of an embodiment, and here is omitted.Specifically：
Identification module described second to second test data using arest neighbors sorting algorithm specifically for training number
According to being identified on collection.
The embodiment of the invention discloses a kind of specific image-recognizing method analyzed based on border Fisher.Specifically：
The present invention of the embodiment of the present invention is tested on Jaffe face expression databases, and the data set is expression recognition
One of most popular data set in research.Jaffe face databases include 213 expression pictures of 10 Japanese womens, altogether
Containing 7 kinds of basic facial expressions, be respectively it is angry, glad, sad, surprised, detest, it is frightened and poker-faced.In this example, use
Eyes of the manual method positioning per pictures, the segmentation of expression picture is carried out according to the position of eyes, the face of every pictures is obtained
Portion expression region.Every human face expression figure is sized as 32*32.
S501, obtains the first training datasetWherein xi∈RD, yi∈ { 1,2 ..., c }, N is sample
Number, N=140, D is the dimension D=1024, y of sampleiIt is xiClass label, c be classification sum, c=7；
S502, the object function to finding projection matrixSolve, obtain the projection square
Battle array；Wherein, projection matrix is P and P ∈ RD×m, m is the subspace size of setting；X∈RD×N, it is first training dataset
In the sample matrix made of training data；Lb∈RN×N, it is Laplacian Matrix between class；Lw∈RN×N, it is Laplce in class
Matrix；A is to matrix X (Lw+Lb)XTThe matrix being made up of after feature decomposition the corresponding characteristic vector of its nonzero eigenvalue.
Therefore, the projection matrix can maximum retention data geometry.
S503, is solved using linear Bregman alternative mannersTo the projection matrix P rarefactions, obtain
To the sparse projection matrix V, wherein | | V | |1It is 1 norm.
Specifically, by projection matrix rarefaction, solved using linear Bregman alternative mannersTo obtain
Sparse projection matrix V after projection matrix rarefaction.
The weight of LS-SVM sparseness, control original variable and feature is carried out by linear Bregman iteration, can be obtained more
Generalization ability and discriminating power well.
S504, first training datasetAccording to the sparse projection matrix V project obtaining second
Training datasetWherein zi=VTxi。
S505, receives the first test data of images to be recognized, the sparse projection is passed through to first test data
Matrix project obtaining the second test data；
Specifically, the first test data x, wherein x ∈ R of figure to be identified are receivedD, LS-SVM sparseness is carried out to x, that is,
Say and projected x by sparse projection matrix V, obtain the second test data z, wherein z ∈ Rm。
S506, second test data is identified on second training dataset using sorting algorithm.
Specifically, using sorting algorithm to the second test data z in the second training datasetIt is upper to be classified,
Test data is identified.
The embodiment of the present invention is chosen 20 face samples from every class of Jaffe databases and tested, and 7 classes expression is altogether
140 training samples, remaining all sample of data set are used as test sample.Experiment is repeated 10 times, 10 recognition results of record
Average value, control methods is the MFA algorithms and the MFA after PCA dimensionality reductions for not carrying out rarefaction.Experimental result is as shown in table 1,
Even if from experimental result as can be seen that a kind of image-recognizing method based on border Fisher analyses of the invention provided is unmated
Also preferable recognition effect can be possessed in the case of PCA dimensionality reductions, discrimination can be effectively lifted.
Table 1：Expression Recognition rate (%) based on sparse border Fisher analyses and border Fisher analyses
| Method | Border Fisher is analyzed | PCA+ borders Fisher is analyzed | The present invention |
| Average recognition rate | 46.44±4.68 | 78.63±4.67 | 79.04±3.42 |
The embodiment of each in this specification is described by the way of progressive, and what each embodiment was stressed is and other
Between the difference of embodiment, each embodiment identical similar portion mutually referring to.
The foregoing description of the disclosed embodiments, enables professional and technical personnel in the field to realize or using the present invention.
A variety of modifications to these embodiments will be apparent for those skilled in the art, as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, it is of the invention
The embodiments shown herein is not intended to be limited to, and is to fit to and principles disclosed herein and features of novelty phase one
The most wide scope caused.
Claims (10)
1. a kind of image-recognizing method based on sparse border Fisher algorithms, it is characterised in that including:
The first training dataset is obtained, projection matrix is obtained according to first training dataset；
LS-SVM sparseness is carried out to the projection matrix, sparse projection matrix is obtained；
First training dataset is projected by the sparse projection matrix, the second training dataset is obtained；
The first test data of images to be recognized is received, first test data is thrown by the sparse projection matrix
Shadow obtains the second test data；
Second test data is identified on second training dataset using sorting algorithm.
2. image-recognizing method according to claim 1, it is characterised in that the utilization sorting algorithm is surveyed to described second
Examination data are identified on second training dataset, including：
Second test data is identified on second training dataset using arest neighbors sorting algorithm.
3. image-recognizing method according to claim 1, it is characterised in that the training dataset of acquisition first, according to
First training dataset obtains projection matrix, including：
Obtain the first training datasetWherein xi∈RD, yi∈ { 1,2 ..., c }, N are the numbers of sample, and D is sample
This dimension, yiIt is xiClass label, c be classification sum；
Object function to finding projection matrixSolve, obtain the projection matrix；Wherein, throw
Shadow matrix is P and P ∈ RD×m, m is the subspace size of setting；X∈RD×N, it is the training number that first training data is concentrated
According to the sample matrix made；Lb∈RN×N, it is Laplacian Matrix between class；Lw∈RN×N, it is Laplacian Matrix in class；A is pair
Matrix X (Lw+Lb)XTThe matrix being made up of after feature decomposition the corresponding characteristic vector of its nonzero eigenvalue.
4. image-recognizing method according to claim 3, it is characterised in that carried out to the projection matrix at rarefaction
Reason, obtains sparse projection matrix, including：
Solved using linear Bregman alternative mannersTo the projection matrix P rarefactions, obtain described sparse
Projection matrix V, wherein | | V | |1It is 1 norm.
5. image-recognizing method according to claim 4, it is characterised in that described to pass through first training dataset
The sparse projection matrix is projected, and obtains the second training dataset, including：
First training datasetAccording to the sparse projection matrix V project obtaining the second training datasetWherein zi=VTxi。
6. a kind of pattern recognition device based on sparse border Fisher algorithms, it is characterised in that including：
Projection matrix acquisition module, for obtaining the first training dataset, obtains projecting square according to first training dataset
Battle array；
LS-SVM sparseness module, for carrying out LS-SVM sparseness to the projection matrix, obtains sparse projection matrix；
Training dataset projection module, for first training dataset to be projected by the sparse projection matrix,
Obtain the second training dataset；
Test data projection module, the first test data for receiving images to be recognized, passes through to first test data
The sparse projection matrix project obtaining the second test data；
Identification module, for being known using sorting algorithm to second test data on second training dataset
Not.
7. pattern recognition device according to claim 6, it is characterised in that the identification module is specifically for using recently
Second test data is identified adjacent sorting algorithm on second training dataset.
8. pattern recognition device according to claim 6, it is characterised in that the projection matrix acquisition module, including：
First training dataset acquiring unit, for obtaining the first training datasetWherein xi∈RD, yi∈{1,
2 ..., c }, N is the number of sample, and D is the dimension of sample, yiIt is xiClass label, c be classification sum；
Projection matrix acquiring unit, for the object function to finding projection matrixSolve, obtain
The projection matrix；Wherein, projection matrix is P and P ∈ RD×m, m is the subspace size of setting；X∈RD×N, it is described first
The sample matrix that the training data that training data is concentrated is made；Lb∈RN×N, it is Laplacian Matrix between class；Lw∈RN×N, it is class
Interior Laplacian Matrix；A is to matrix X (Lw+Lb)XTIt is made up of after feature decomposition the corresponding characteristic vector of its nonzero eigenvalue
Matrix.
9. pattern recognition device according to claim 8, it is characterised in that the LS-SVM sparseness module, specifically for
Solved using linear Bregman alternative mannersTo the projection matrix P rarefactions, the sparse projection is obtained
Matrix V, wherein | | V | |1It is 1 norm.
10. pattern recognition device according to claim 9, it is characterised in that the training dataset projection module, specifically
For first training datasetAccording to the sparse projection matrix V project obtaining the second training data
CollectionWherein zi=VTxi。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710425040.4A CN107203786A (en) | 2017-06-06 | 2017-06-06 | Image-recognizing method and device based on sparse border Fisher algorithms |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710425040.4A CN107203786A (en) | 2017-06-06 | 2017-06-06 | Image-recognizing method and device based on sparse border Fisher algorithms |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107203786A true CN107203786A (en) | 2017-09-26 |
Family
ID=59908399
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710425040.4A Pending CN107203786A (en) | 2017-06-06 | 2017-06-06 | Image-recognizing method and device based on sparse border Fisher algorithms |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107203786A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110852304A (en) * | 2019-11-22 | 2020-02-28 | 重庆大学 | Hyperspectral data processing method based on deep learning method |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101021897A (en) * | 2006-12-27 | 2007-08-22 | 中山大学 | Two-dimensional linear discrimination human face analysis identificating method based on interblock correlation |
| CN103955676A (en) * | 2014-05-12 | 2014-07-30 | 苏州大学 | Human face identification method and system |
| CN104537384A (en) * | 2015-01-21 | 2015-04-22 | 西安电子科技大学 | SAR (synthetic aperture radar) target identification method combined with likelihood ratio decision |
| CN104616000A (en) * | 2015-02-27 | 2015-05-13 | 苏州大学 | Human face recognition method and apparatus |
| CN104794498A (en) * | 2015-05-07 | 2015-07-22 | 西安电子科技大学 | Image classification method based on combination of SRC and MFA |
| CN104899578A (en) * | 2015-06-26 | 2015-09-09 | 苏州大学张家港工业技术研究院 | Method and device for face identification |
-
2017
- 2017-06-06 CN CN201710425040.4A patent/CN107203786A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101021897A (en) * | 2006-12-27 | 2007-08-22 | 中山大学 | Two-dimensional linear discrimination human face analysis identificating method based on interblock correlation |
| CN103955676A (en) * | 2014-05-12 | 2014-07-30 | 苏州大学 | Human face identification method and system |
| CN104537384A (en) * | 2015-01-21 | 2015-04-22 | 西安电子科技大学 | SAR (synthetic aperture radar) target identification method combined with likelihood ratio decision |
| CN104616000A (en) * | 2015-02-27 | 2015-05-13 | 苏州大学 | Human face recognition method and apparatus |
| CN104794498A (en) * | 2015-05-07 | 2015-07-22 | 西安电子科技大学 | Image classification method based on combination of SRC and MFA |
| CN104899578A (en) * | 2015-06-26 | 2015-09-09 | 苏州大学张家港工业技术研究院 | Method and device for face identification |
Non-Patent Citations (4)
| Title |
|---|
| QIAO Z 等: "Sparse Linear Discriminant Analysis with Applications to High Dimensional Low Sample Size Data", 《INTERNATIONAL JOURNAL OF APPLIED MATHEMATICS》 * |
| QINGFENG LIU 等: "Sparse Representation Based Complete Kernel Marginal Fisher Analysis Framework for Computational Art Painting Categorization", 《COMPUTER VISION–ECCV 2016》 * |
| WANG, Z 等: "Facial expression recognition using sparse local Fisher discriminant analysis", 《NEUROCOMPUTING》 * |
| 李峰 等: "半监督的稀疏保持二维边界Fisher分析降维算法", 《计算机辅助设计与图形学学报》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110852304A (en) * | 2019-11-22 | 2020-02-28 | 重庆大学 | Hyperspectral data processing method based on deep learning method |
| CN110852304B (en) * | 2019-11-22 | 2022-03-22 | 重庆大学 | Hyperspectral data processing method based on deep learning method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108492343B (en) | Image synthesis method for training data for expanding target recognition | |
| CN104143079B (en) | The method and system of face character identification | |
| CN103164694B (en) | A kind of human action knows method for distinguishing | |
| CN106203356B (en) | A kind of face identification method based on convolutional network feature extraction | |
| CN108491077A (en) | A kind of surface electromyogram signal gesture identification method for convolutional neural networks of being divided and ruled based on multithread | |
| CN108171133B (en) | Dynamic gesture recognition method based on characteristic covariance matrix | |
| CN105205449B (en) | Sign Language Recognition Method based on deep learning | |
| CN108960059A (en) | A kind of video actions recognition methods and device | |
| Osherov et al. | Increasing cnn robustness to occlusions by reducing filter support | |
| CN107273876A (en) | A kind of micro- expression automatic identifying method of ' the grand micro- transformation models of to ' based on deep learning | |
| CN108898620A (en) | Method for tracking target based on multiple twin neural network and regional nerve network | |
| CN105678231A (en) | Pedestrian image detection method based on sparse coding and neural network | |
| CN105005765A (en) | Facial expression identification method based on Gabor wavelet and gray-level co-occurrence matrix | |
| CN104598889B (en) | The method and apparatus of Human bodys' response | |
| CN109410168A (en) | For determining the modeling method of the convolutional neural networks model of the classification of the subgraph block in image | |
| CN105046219A (en) | Face identification system | |
| CN104143076B (en) | The matching process of face shape and system | |
| CN109034099A (en) | A kind of expression recognition method and device | |
| CN105654141A (en) | Isomap and SVM algorithm-based overlooked herded pig individual recognition method | |
| CN104063721B (en) | A kind of human behavior recognition methods learnt automatically based on semantic feature with screening | |
| Zhou et al. | Multi-classification of skin diseases for dermoscopy images using deep learning | |
| CN107808376A (en) | A kind of detection method of raising one's hand based on deep learning | |
| CN109685045A (en) | A kind of Moving Targets Based on Video Streams tracking and system | |
| Zhou et al. | Sampling-attention deep learning network with transfer learning for large-scale urban point cloud semantic segmentation | |
| CN107392251A (en) | A kind of method that target detection network performance is lifted using category images |
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: 20170926 |