CN110458137A - A kind of diameter radar image recognition methods based on two-stage multi-task learning - Google Patents
A kind of diameter radar image recognition methods based on two-stage multi-task learning Download PDFInfo
- Publication number
- CN110458137A CN110458137A CN201910765203.2A CN201910765203A CN110458137A CN 110458137 A CN110458137 A CN 110458137A CN 201910765203 A CN201910765203 A CN 201910765203A CN 110458137 A CN110458137 A CN 110458137A
- Authority
- CN
- China
- Prior art keywords
- feature
- heterogeneous
- target image
- image
- training
- 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
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- 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
- G06F18/2136—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on sparsity criteria, e.g. with an overcomplete basis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of diameter radar image recognition methods based on two-stage multi-task learning, comprising: extracts the heterogeneous feature of multiclass of target image and training image collection;The linear combination of the corresponding heterogeneous feature of all training images is concentrated to indicate with training image every heterogeneous feature of one kind of target image;Extract the coefficient vector of every heterogeneous feature of one kind;Obtain the M neighbour part subset and the new dictionary of composition of every heterogeneous feature of one kind;Obtaining collaboration indicates coefficient matrix;The recognition result of target image is obtained into criterion using minimal reconstruction error.First stage of the invention obtains best local subset of the target image in training sample, and accordingly update the dictionary of each feature, greatly reduce the interference of the training sample far from target image, it is interfered so as to avoid caused by exceptional value, classification performance is substantially better than the study of multitask rarefaction representation and multitask coordinated expression learning algorithm, and has robustness to extensive regularization parameter.
Description
Technical field
The present invention relates to Technology of Radar Target Identification fields, and in particular to a kind of synthesis hole based on two-stage multi-task learning
Aperture radar image recognition methods.
Background technique
Radar image target identification is the important subject in science and techniques of defence field.It is thunder using the method for indicating study
Up to approach more effective in image domains, key technology is exactly to indicate the foundation of learning model.
The common radar image target identification method based on expression study has: (1) based on sparse representation method, indicating
Under the sparse constraint of coefficient, obtain uniquely indicating coefficient solution, what rarefaction representation utilized is " competition " relationship between sample,
The method that algorithm complexity higher (2) is indicated based on collaboration, collaboration indicate to utilize " cooperation " relationship between sample, actually L2
Optimization problem under norm constraint greatly reduces computation complexity, but need in the accuracy rate of algorithm and robustness into one
Step improves (3) based on multi-task learning and the method for indicating that study is combined, and this method has become at present indicates the another of study
Development trend.Multi-task learning takes full advantage of that identification mission is similar enough or this related characteristic to a certain extent, favorably
In the generalization ability for improving recognizer.The a variety of different types of features for extracting image see the expression identification of every category feature
Work is a task, and establishing multiple features combining indicates model.Multi-task learning combine indicate study method image recognition with
And there is very extensive utilization in class object identification field, can excavate the shared training sample mode between different task, from
And ensure to select correct training sample, reduce the interference of error category training sample.
The present invention proposes a kind of base on the basis of indicating the method for learning to be combined based on multi-task learning existing
In the diameter radar image recognition methods of two-stage multi-task learning, it is more to indicate that learning method is utilized based on two-stage multitask
Feature cooperates with discriminating power, and rarefaction representation is combined with the expression learning ability that collaboration indicates, utilizes in the first stage
L2,1The multitask rarefaction representation of norm regularization obtains best local subset of the target image in training sample, and accordingly more
The dictionary of new each feature, greatly reduces the interference of the training sample far from target image, causes so as to avoid exceptional value
Interference, classification performance is substantially better than the study of multitask rarefaction representation and multitask coordinated expression learning algorithm, and to extensive
Regularization parameter has robustness.
Summary of the invention
In view of the above shortcomings of the prior art, the present invention is in the existing side combined based on multi-task learning with expression study
A kind of diameter radar image recognition methods based on two-stage multi-task learning is proposed on the basis of method, is based on two-stage more
Business indicates that learning method is utilized multiple features and cooperates with discriminating power, and by rarefaction representation and cooperates with the expression learning ability indicated
It combines, utilizes L in the first stage2,1The multitask rarefaction representation of norm regularization obtains target image in training sample
Best part subset, and the dictionary of each feature is accordingly updated, the interference of the training sample far from target image is greatly reduced,
It is interfered so as to avoid caused by exceptional value, classification performance is substantially better than the study of multitask rarefaction representation and multitask coordinated expression
Learning algorithm, and there is robustness to extensive regularization parameter.
Present invention employs the following technical solutions:
A kind of diameter radar image recognition methods based on two-stage multi-task learning, comprising:
S1, target image and training image collection are obtained;
S2, the heterogeneous feature of multiclass for extracting the target image and the training image collection;
S3, concentrate all training images corresponding with the training image every heterogeneous feature of one kind of the target image
The linear combination of heterogeneous feature indicates;
S4, the linear combination based on the heterogeneous feature indicate to extract the coefficient vector of every heterogeneous feature of one kind;
S5, the coefficient vector based on every heterogeneous feature of one kind obtain the M neighbour part of every heterogeneous feature of one kind
Subset and the new dictionary of composition, M neighbour part subset are the proper subclass of coefficient vector;
S6, new dictionary is formed to indicate target image based on M neighbour part subset, obtaining collaboration indicates coefficient matrix;
S7, coefficient matrix is indicated based on collaboration, obtains the recognition result of target image into criterion using minimal reconstruction error.
Preferably, the heterogeneous feature of the multiclass includes any one in PCA feature, Wavelet Transform Feature and 2DSZM feature
Kind is a variety of.
Preferably, in step S3:
……
In formula, ykFor the heterogeneous feature of kth class of target image, P is heterogeneous feature class number,For ykCorresponding i-th of instruction
Practicing the coefficient vector atom of image, n is the number that the training image concentrates training image,For ykCorresponding i-th of training
The heterogeneous feature of image;
By the heterogeneous character representation of target image are as follows:
y1=X1a'1
……
yk=Xka'k
yP=XPa'P
In formula, XkFor ykCorresponding heterogeneous feature vector, a'kFor ykCorresponding expression coefficient vector.
Preferably, in step S4:
Establish L2,1The multitask sparse representation model of norm regularization:
In formula, η is balance parameters, | | | |2,1Indicate L2,1Norm, A are coefficient vector collection, A=[a1…aP, | | | |F
The F norm of representing matrix.
Preferably, in step S5:
Maximum preceding Q coefficient vector atom is extracted from the coefficient vector of every heterogeneous feature of one kind and duplicate removal obtains atom
Collection,The k pattern feature vector of j-th of atom is concentrated for atom;
Based on formulaCalculate distance measure amount ej, take corresponding ejThe smallest preceding M atom forms M
Neighbour part subset;
0 is set by the coefficient vector atom other than atom in the subset of M neighbour part, obtains new heterogeneous feature vector simultaneously
Using new heterogeneous feature vector as new dictionary, X'kFor XkCorresponding new dictionary.
Preferably, in step S6:
New dictionary is formed based on M neighbour part subset to indicate target image:
In formula, ρ is balance parameters, | | | |FThe F norm of representing matrix, B are that collaboration indicates coefficient matrix, B=[b1…
bP], bkFor ykCorresponding new expression coefficient vector.
In conclusion the invention discloses a kind of diameter radar image identification sides based on two-stage multi-task learning
Method, comprising: S1, obtain target image and training image collection;S2, the multiclass for extracting the target image and the training image collection
Heterogeneous feature;S3, concentrate all training images corresponding with the training image every heterogeneous feature of one kind of the target image
Heterogeneous feature linear combination indicate;S4, the linear combination based on the heterogeneous feature indicate to extract every heterogeneous feature of one kind
Coefficient vector;S5, the coefficient vector based on every heterogeneous feature of one kind obtain the M neighbour office of every heterogeneous feature of one kind
Portion's subset and the new dictionary of composition, M neighbour part subset are the proper subclass of coefficient vector;S6, it is formed newly based on M neighbour part subset
Dictionary indicates target image, and obtaining collaboration indicates coefficient matrix;S7, coefficient matrix is indicated based on collaboration, utilizes minimal reconstruction
Error obtains the recognition result of target image into criterion.On the basis of existing technology, the present invention is based on two-stage multitask expressions
Multiple features collaboration discriminating power is utilized in learning method, and rarefaction representation is mutually tied with the expression learning ability indicated is cooperateed with
It closes, utilizes L in the first stage2,1It is best in training sample that the multitask rarefaction representation of norm regularization obtains target image
Local subset, and the dictionary of each feature is accordingly updated, the interference of the training sample far from target image is greatly reduced, thus
Interference caused by exceptional value is avoided, classification performance is substantially better than the study of multitask rarefaction representation and multitask coordinated expression study
Algorithm, and there is robustness to extensive regularization parameter.
Detailed description of the invention
In order to keep the purposes, technical schemes and advantages of invention clearer, the present invention is made into one below in conjunction with attached drawing
The detailed description of step, in which:
Fig. 1 is a kind of stream of the diameter radar image recognition methods based on two-stage multi-task learning disclosed by the invention
Cheng Tu;
Fig. 2 is a certain radar target image;
Fig. 3 is that the second stage collaboration of PCA feature indicates coefficient;
Fig. 4 is that the second stage collaboration of wavelet character indicates coefficient;
Fig. 5 is that the second stage collaboration of 2DSZM feature indicates coefficient;
Fig. 6 is the reconstructed error that three category features in learning method are indicated based on two-stage multitask;
Fig. 7 is to indicate that learning method, direct multitask sparse representation method and direct multitask are assisted based on two-stage multitask
The curve graph changed with the discrimination of three kinds of methods of representation method with intrinsic dimensionality;
Fig. 8 is to indicate that learning method, direct multitask sparse representation method and direct multitask are assisted based on two-stage multitask
With three kinds of method discriminations of representation method with the curve graph for anywhere rule changing Parameters variation;
Fig. 9 is the identification indicated in learning method for two stage Different Rule Parameters variations based on two-stage multitask
Effect picture.
Specific embodiment
The present invention is described in further detail with reference to the accompanying drawing.
As shown in Figure 1, the invention discloses a kind of diameter radar image identification sides based on two-stage multi-task learning
Method, comprising:
S1, target image and training image collection are obtained;
In the present invention, the corresponding label of training image can be also obtained during obtaining training image collection, for subsequent
Identification.
S2, the heterogeneous feature of multiclass for extracting the target image and the training image collection;
It is well known that each category feature is all to describe image from some angle, therefore it cannot be distinguished comprising all
Other information.In addition, it is unpractiaca for extracting an optimal characteristics.Therefore, for image identification system, better method is to combine
Multiclass feature is identified, rather than only uses single features.
S3, concentrate all training images corresponding with the training image every heterogeneous feature of one kind of the target image
The linear combination of heterogeneous feature indicates;
S4, the linear combination based on the heterogeneous feature indicate to extract the coefficient vector of every heterogeneous feature of one kind;
S5, the coefficient vector based on every heterogeneous feature of one kind obtain the M neighbour part of every heterogeneous feature of one kind
Subset and the new dictionary of composition, M neighbour part subset are the proper subclass of coefficient vector;
S6, new dictionary is formed to indicate target image based on M neighbour part subset, obtaining collaboration indicates coefficient matrix;
S7, coefficient matrix is indicated based on collaboration, obtains the recognition result of target image into criterion using minimal reconstruction error.
The recognition result of target image is obtained for the prior art, herein not into criterion using minimal reconstruction error in the present invention
It repeats again.
First stage of the invention is exactly the individual features that every category feature of target image is expressed as to all training samples
Linear combination, and determine using the multitask rarefaction representation ability of training sample the M neighbour part subset of target image.
This is because in principle, current target image and its some neighbour's sample standard deviations should be from same class, that is to say, that
These neighbour's samples are maximum to the contribution of recognition target image.Therefore, the first step of algorithm is exactly to detect far from target image
Those of training sample, and assume that these samples do not influence in categorised decision.This precise classification for target image
It is helpful.With local training sample, rather than all training samples carry out recognition target image, can substantially reduce far from mesh
The interference of those of logo image training sample.The second stage of algorithm is exactly will be using multitask coordinated expression Learning Principle, benefit
Target image is indicated with the new dictionary that M neighbour part subset forms, and carries out images steganalysis using expression result.The
It is to be that second level, which uses the reason of multitask coordinated expression: it is multitask coordinated to indicate simple, and have closed solutions.The side proposed
Method has following reason: its first stage has determined some with the most like training sample of current target image, and will be identified
Candidate categories label of the class label of training sample as target image.Since the class label of the training sample identified is usually
The a subset of all training samples determines finally so final classification becomes in a candidate class label from lesser amt
The problem of class label.When the class label of real target image is one in those training samples for determining, this will be to the
It is very helpful that two-stage carries out accurately classification.This algorithm proposed by the present invention has not only sufficiently excavated the table of multiclass feature
Show learning ability;And the interference of those unrelated dictionary atoms is greatly reduced, therefore effectively improve recognition performance.
When it is implemented, the heterogeneous feature of multiclass includes appointing in PCA feature, Wavelet Transform Feature and 2DSZM feature
It anticipates one or more.
In present invention, it is preferable to simultaneously using PCA feature, Wavelet Transform Feature and 2DSZM feature these three features, these three
It is characterized in the SAR image feature from different feature extraction angle extractions, has preferable complementary.
When it is implemented, in step S3:
……
In formula, ykFor the heterogeneous feature of kth class of target image, P is heterogeneous feature class number,For ykCorresponding i-th of instruction
Practicing the coefficient vector atom of image, n is the number that the training image concentrates training image,For ykCorresponding i-th of training
The heterogeneous feature of image;
By the heterogeneous character representation of target image are as follows:
y1=X1a'1
……
yk=Xka'k
yP=XPa'P
In formula, XkFor ykCorresponding heterogeneous feature vector, a'kFor ykCorresponding expression coefficient vector.
When it is implemented, in step S4:
Establish L2,1The multitask sparse representation model of norm regularization:
In formula, η is balance parameters, | | | |2,1Indicate L2,1Norm, A are coefficient vector collection, A=[a1…aP], | | | |F
The F norm of representing matrix.
According to rarefaction representation principle, the feature from identical training sample should be shared between every kind of feature of target image.
And in view of brings errors such as noise factors, L can will be obtained2,1The multitask sparse representation model of norm regularization.More
Regularization term has selected the L of A in business sparse representation model2,1Norm regularization term.This is because for coming from same target figure
For the different characteristic of picture, the position of nonzero coefficient should be similar in all sparse vectors, i.e., dilute between shared class
The mode of dredging.And due to the difference between different types of feature, the shared corresponding coefficient values of training sample of these in sparse vector
It is different.Under this hypothesis, indicate that nonzero coefficient should be in Xiang Tonghang, a joint sparse regularization L in A2,1It can be set to
In the non-zero row of expression for selecting smallest number on A.Multitask sparse representation model can be solved by APG algorithm.
When it is implemented, in step S5:
Maximum preceding Q coefficient vector atom is extracted from the coefficient vector of every heterogeneous feature of one kind and duplicate removal obtains atom
Collection,The k pattern feature vector of j-th of atom is concentrated for atom;
Based on formulaCalculate distance measure amount ej, take corresponding ejThe smallest preceding M atom forms M
Neighbour part subset;
0 is set by the coefficient vector atom other than atom in the subset of M neighbour part, obtains new heterogeneous feature vector simultaneously
Using new heterogeneous feature vector as new dictionary, X'kFor XkCorresponding new dictionary.
By multitask sparse representation model, available A, it is each column be corresponding different characteristic expression coefficient to
Amount.When indicating the feature of target image, the individual features of each training sample can make the contribution of oneself.Substantially, A
Each column be rarefaction representation coefficient vector of the target image under character pair.The contribution of i-th of training sample in dictionary
It can be assessed by the size of corresponding rarefaction representation coefficient value.One big expression coefficient means that i-th of training sample
There is very big contribution to expression target image.Due to there is the multiple rarefaction representation coefficient vectors for corresponding to multiple features;Therefore we
M neighbour part subset is obtained using above-mentioned method.In this way, the dictionary X that the feature of every one kind is constitutedkIt is all accordingly updated to new
Dictionary X'k.In this way, can greatly reduce uncorrelated atom pair indicates study bring interference, to reduce erroneous judgement.
When it is implemented, in step S6:
New dictionary is formed based on M neighbour part subset to indicate target image:
In formula, ρ is balance parameters, | | | |FThe F norm of representing matrix, B are that collaboration indicates coefficient matrix, B=[b1…
bP], bkFor ykCorresponding new expression coefficient vector.
It is multitask coordinated to indicate that there are analytic solutions.That is:
It is to be calculated simply because it has, it is possible to reduce the advantage of computation burden using multitask coordinated expression.
After the multitask coordinated expression in the second level, target type is determined according to multitask reconstructed error minimum, it is as follows
Formula:
In order to prove technical effect of the invention, we are using radar target image as shown in Figure 2 as target image.
For given radar target image, the heterogeneous feature of multiclass of all training samples and target image, including PCA spy are extracted first
Sign, Wavelet Transform Feature, 2DSZM feature.Multitask in the first stage indicates to utilize L in study2,1Norm regularization it is more
Task sparse representation model acquires the coefficient vector of every category feature.It is maximum to select Q coefficient value before finding in each coefficient vector
Atom, remove duplicate atom, by these atoms form a new set, further according to the Europe of characteristic vector and target image
M atom forms final M neighbour part subset before formula distance is selected.It will be corresponding in the training sample dictionary of each feature
M training sample retains, and other dictionary atoms are set to 0 update dictionary.Second stage utilizes multitask coordinated expression, obtains
The collaboration of the feature of tertiary target image indicates the reconstructed error of coefficient (as shown in figure 3, figure 4 and figure 5) and three category features (such as
Fig. 6).Fig. 7 is that study and direct multitask sparse representation method and direct multitask coordinated method are indicated based on two-stage multitask
With the recognition effect that intrinsic dimensionality changes, which illustrates that two-stage multi-task learning method still can be under dimension situation of change
Keep more superior recognition performance.Fig. 8 is that (method of the invention is TSMRL to three kinds of methods, and in addition two methods are respectively
MSRC and MCRC) anywhere rule change the recognition performance curve of Parameters variation, which has absolutely proved this method for regularisation parameter
Value has preferable robustness, and good recognition effect (this hair can be obtained in the very big variation range of regularisation parameter value
Two kinds of regularisation parameters, λ and ρ can be used in bright).Fig. 9 is two-stage multi-task learning method for two stage Different Rules
The recognition effect figure of Parameters variation.These experimental results support the superiority of method proposed by the present invention.
Finally, it is stated that the above examples are only used to illustrate the technical scheme of the present invention and are not limiting, although passing through ginseng
According to the preferred embodiment of the present invention, invention has been described, it should be appreciated by those of ordinary skill in the art that can
To make various changes to it in the form and details, without departing from the present invention defined by the appended claims
Spirit and scope.
Claims (6)
1. a kind of diameter radar image recognition methods based on two-stage multi-task learning characterized by comprising
S1, target image and training image collection are obtained;
S2, the heterogeneous feature of multiclass for extracting the target image and the training image collection;
S3, concentrate all training images corresponding heterogeneous with the training image every heterogeneous feature of one kind of the target image
The linear combination of feature indicates;
S4, the linear combination based on the heterogeneous feature indicate to extract the coefficient vector of every heterogeneous feature of one kind;
S5, the coefficient vector based on every heterogeneous feature of one kind obtain the M neighbour part subset of every heterogeneous feature of one kind
And new dictionary is formed, M neighbour part subset is the proper subclass of coefficient vector;
S6, new dictionary is formed to indicate target image based on M neighbour part subset, obtaining collaboration indicates coefficient matrix;
S7, coefficient matrix is indicated based on collaboration, obtains the recognition result of target image into criterion using minimal reconstruction error.
2. as described in claim 1 based on the diameter radar image recognition methods of two-stage multi-task learning, feature exists
In the heterogeneous feature of multiclass includes any one or more in PCA feature, Wavelet Transform Feature and 2DSZM feature.
3. as described in claim 1 based on the diameter radar image recognition methods of two-stage multi-task learning, feature exists
In in step S3:
……
In formula, ykFor the heterogeneous feature of kth class of target image, P is heterogeneous feature class number,For ykCorresponding i-th of training image
Coefficient vector atom, n be the training image concentrate training image number,For ykCorresponding i-th of training image
Heterogeneous feature;
By the heterogeneous character representation of target image are as follows:
y1=X1a'1
……
yk=Xka'k
yP=XPa'P
In formula, XkFor ykCorresponding heterogeneous feature vector, a'kFor ykCorresponding expression coefficient vector.
4. as claimed in claim 3 based on the diameter radar image recognition methods of two-stage multi-task learning, feature exists
In in step S4:
Establish L2,1The multitask sparse representation model of norm regularization:
In formula, η is balance parameters, | | | |2,1Indicate L2,1Norm, A are coefficient vector collection, A=[a1...aP], | | | |FTable
Show the F norm of matrix.
5. as claimed in claim 4 based on the diameter radar image recognition methods of two-stage multi-task learning, feature exists
In in step S5:
Maximum preceding Q coefficient vector atom is extracted from every heterogeneous characteristic quantity of one kind and duplicate removal obtains atom collection,For atom
Concentrate the k pattern feature vector of j-th of atom;
Based on formulaCalculate distance measure amount ej, take corresponding ejThe smallest preceding M atom forms M neighbour
Local subset;
0 is set by the coefficient vector atom other than atom in the subset of M neighbour part, obtains new heterogeneous feature vector and will be new
Heterogeneous feature vector as new dictionary, X'kFor XkCorresponding new dictionary.
6. as claimed in claim 5 based on the diameter radar image recognition methods of two-stage multi-task learning, feature exists
In in step S6:
New dictionary is formed based on M neighbour part subset to indicate target image:
In formula, ρ is balance parameters, | | | |FThe F norm of representing matrix, B are that collaboration indicates coefficient matrix, B=[b1...bP], bk
For ykCorresponding new expression coefficient vector.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910765203.2A CN110458137A (en) | 2019-08-19 | 2019-08-19 | A kind of diameter radar image recognition methods based on two-stage multi-task learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910765203.2A CN110458137A (en) | 2019-08-19 | 2019-08-19 | A kind of diameter radar image recognition methods based on two-stage multi-task learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110458137A true CN110458137A (en) | 2019-11-15 |
Family
ID=68487723
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910765203.2A Pending CN110458137A (en) | 2019-08-19 | 2019-08-19 | A kind of diameter radar image recognition methods based on two-stage multi-task learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110458137A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
FR3115378A1 (en) * | 2020-10-20 | 2022-04-22 | Biomerieux | Method for classifying an input image representing a particle in a sample |
CN115470873A (en) * | 2022-11-14 | 2022-12-13 | 中国人民解放军国防科技大学 | Radar radiation source identification method and system |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104899549A (en) * | 2015-04-17 | 2015-09-09 | 重庆大学 | SAR target recognition method based on range profile time-frequency image identification dictionary learning |
CN106096506A (en) * | 2016-05-28 | 2016-11-09 | 重庆大学 | Based on the SAR target identification method differentiating doubledictionary between subclass class |
-
2019
- 2019-08-19 CN CN201910765203.2A patent/CN110458137A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104899549A (en) * | 2015-04-17 | 2015-09-09 | 重庆大学 | SAR target recognition method based on range profile time-frequency image identification dictionary learning |
CN106096506A (en) * | 2016-05-28 | 2016-11-09 | 重庆大学 | Based on the SAR target identification method differentiating doubledictionary between subclass class |
Non-Patent Citations (1)
Title |
---|
XINZHENG ZHANG 等: "Two-Stage Multi-Task Representation Learning for Synthetic Aperture Radar (SAR) Target Images Classification", 《SENSORS》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
FR3115378A1 (en) * | 2020-10-20 | 2022-04-22 | Biomerieux | Method for classifying an input image representing a particle in a sample |
WO2022084617A1 (en) * | 2020-10-20 | 2022-04-28 | Biomerieux | Method for classifying an input image containing a particle in a sample |
CN115470873A (en) * | 2022-11-14 | 2022-12-13 | 中国人民解放军国防科技大学 | Radar radiation source identification method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Song et al. | Region-based quality estimation network for large-scale person re-identification | |
CN108898180B (en) | Depth clustering method for single-particle cryoelectron microscope images | |
CN104463250B (en) | A kind of Sign Language Recognition interpretation method based on Davinci technology | |
CN103400388B (en) | A kind of method utilizing RANSAC to eliminate Brisk key point error matching points pair | |
CN105184260B (en) | A kind of image characteristic extracting method and pedestrian detection method and device | |
CN104123529B (en) | human hand detection method and system | |
CN107346550B (en) | It is a kind of for the three dimensional point cloud rapid registering method with colouring information | |
Huber et al. | Mask-invariant face recognition through template-level knowledge distillation | |
CN109102024A (en) | A kind of Layer semantics incorporation model finely identified for object and its implementation | |
CN101996308A (en) | Human face identification method and system and human face model training method and system | |
Ji et al. | Memorizing complementation network for few-shot class-incremental learning | |
CN110458137A (en) | A kind of diameter radar image recognition methods based on two-stage multi-task learning | |
CN110390289A (en) | Based on the video security protection detection method for censuring understanding | |
CN109409201A (en) | A kind of pedestrian's recognition methods again based on shared and peculiar dictionary to combination learning | |
CN110008844A (en) | A kind of long-term gesture tracking method of KCF merging SLIC algorithm | |
CN106203373B (en) | A kind of human face in-vivo detection method based on deep vision bag of words | |
CN110472662A (en) | Image matching algorithm based on improved ORB algorithm | |
CN111461162B (en) | Zero-sample target detection model and establishing method thereof | |
Minvielle et al. | Transfer learning on decision tree with class imbalance | |
CN109697727A (en) | Method for tracking target, system and storage medium based on correlation filtering and metric learning | |
Liu et al. | Double Mask R‐CNN for Pedestrian Detection in a Crowd | |
Chen et al. | Pose-guided spatial alignment and key frame selection for one-shot video-based person re-identification | |
CN105447477B (en) | Formula identification method and device based on formula library | |
CN116704612A (en) | Cross-visual-angle gait recognition method based on contrast domain self-adaptive learning | |
CN110516638A (en) | A kind of sign Language Recognition Method based on track and random forest |
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 |