CN109635738A - A kind of image characteristic extracting method and system - Google Patents
A kind of image characteristic extracting method and system Download PDFInfo
- Publication number
- CN109635738A CN109635738A CN201811525051.0A CN201811525051A CN109635738A CN 109635738 A CN109635738 A CN 109635738A CN 201811525051 A CN201811525051 A CN 201811525051A CN 109635738 A CN109635738 A CN 109635738A
- Authority
- CN
- China
- Prior art keywords
- depth
- multicore
- learning model
- mapping
- basic core
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- 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
- G06F18/24—Classification techniques
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of image characteristic extracting method and systems.This method comprises: building depth multicore maps learning model, each unit of the learning model is a basic core, and each basic core of the learning model interlayer is full connection structure, and each basic core is mutually indepedent in layer;Depth multicore mapping learning model is trained using sample data, the sample data is the multi-dimensional feature data of image;Obtain images to be recognized;Feature extraction is carried out to the images to be recognized using the depth multicore mapping learning model after training.Image characteristic extracting method provided by the invention and system have the characteristics that feature extraction accuracy is high, high-efficient.
Description
Technical field
The present invention relates to image characteristics extraction fields, more particularly to a kind of image characteristic extracting method and system.
Background technique
Research for Remote Sensing Target is all the hot spot of remote sensing fields research all the time, is on the one hand derived from remote sensing figure
It as wide coverage, contains much information, the feature that observation cycle is short, on the other hand also because the various aspects in military and civilian are extensive
Application demand has pushed the development of the application directions such as detection, identification, classification and the analysis of Remote Sensing Target.However along with
The development of imaging technique and target application, Remote Sensing Target research also produce many problems.Firstly, with imaging technique
Development, remote sensing images amount to obtain constantly increases, and the target information covered is more and more, and the redundancy for including is also increasingly
Greatly, the method for high efficiency smart is needed to extract and screen effective information.Secondly, Remote Sensing Target task is also with demand
Complication, fining and become more complicated.For example, script atural object Research Requirements only need to distinguish according to otherness,
Simple division is carried out to remote sensing images, it is now desired to form, develop and profound information carries out intellectual analysis and sentences to it
It is fixed;For another example, script target identification demand only needs to extract object target in the background with basic transformation, need instantly
Will type, trend, situation to target etc. carry out intelligent recognition and analysis, it is therefore desirable to new method extracts target signature.
As the core key that remote sensing target is studied, there is also many problems for feature extraction: conventional method is such as based on frequency domain, small echo exists
Many practical features can be extracted when extracting signal category feature, although very good effect can be obtained in simple image task,
But increasingly complicated requirement is unable to satisfy in complicated analysis task, such as target object similar in shape is (as military in distinguished
Vehicle and common vehicle), conventional method can only obtain target can not but carry out careful differentiation to it, therefore for image and its
The research of feature is particularly important under new image object analysis task demand.Traditional feature extracting method generally uses
The mode of artificial setting feature, and target identification interpretation process is also required to rely on manpower, not only consumes a large amount of manpower in this way
And the time, and recognition accuracy is largely influenced by operator, therefore is extracted using machine autonomous learning
It is characterized in highly effective.For this Type of Nonlinear Object of Remote Sensing Target, have to the non-thread sexuality of feature extraction algorithm
It is certain to require.In numerous algorithms of machine learning, nuclear mapping has extraordinary Nonlinear feature extraction ability, but single
Core, single layer core mapping method have its limitation, and performance is also influenced by selected kernel function type and parameter.
Summary of the invention
The object of the present invention is to provide a kind of image characteristic extracting method and systems, have feature extraction accuracy height, effect
The high feature of rate.
To achieve the above object, the present invention provides following schemes:
A kind of image characteristic extracting method, comprising:
It constructs depth multicore and maps learning model, each unit of the learning model is a basic core,
Practising each basic core of model interlayer is full connection structure, and each basic core is mutually indepedent in layer;
Depth multicore mapping learning model is trained using sample data, the sample data is the more of image
Dimensional feature data;
Obtain images to be recognized;
Feature extraction is carried out to the images to be recognized using the depth multicore mapping learning model after training.
Optionally, described that depth multicore mapping learning model is trained using sample data, it specifically includes:
Set the initial learning rate and initial weight value of the depth multicore mapping learning model;
Input sample data and label into depth multicore mapping learning model, are trained;
Learning model is mapped to the depth multicore according to the loss function of the depth multicore mapping learning model after training
In the weighted value of each basic core be adjusted, until loss function value is less than preset value.
Optionally, the building depth multicore maps learning model, specifically includes:
Determine the type and inner parameter of each basic core;
Building is denoted as the mapping of depth multicore and learns using each basic core as the depth confidence network of the full connection structure of unit
Practise model.
Optionally, the loss function according to the depth multicore mapping learning model after training reflects the depth multicore
The weighted value for penetrating each basis core in learning model is adjusted, and is specifically included:
According toThe weighted value of each basic core in depth multicore mapping learning model is carried out
Adjustment, whereinFor the weighted value of k-th in the t times iterative learning basic core, γkFor the learning rate of k-th of basic core, TSpan
To pass through the loss function for staying a hair to obtain, θkFor the weight of k-th of basic core.
Optionally, the method also includes:
The precision of the depth multicore mapping learning model after training is determined using test data;
Judge whether the precision is less than given threshold;
If it is not, then continuing to be trained depth multicore mapping learning model.
The present invention also provides a kind of image characteristic extraction systems, comprising:
Model construction module, for constructing depth multicore mapping learning model, each unit of the learning model is
One basic core, each basic core of the learning model interlayer are full connection structure, and each basic core is mutually indepedent in layer;
Model training module, it is described for being trained using sample data to depth multicore mapping learning model
Sample data is the multi-dimensional feature data of image;
Image collection module, for obtaining images to be recognized;
Characteristic extracting module, for being carried out using the depth multicore mapping learning model after training to the images to be recognized
Feature extraction.
Optionally, the model training module, specifically includes:
Initialization unit, for setting the initial learning rate and initial weight value of the depth multicore mapping learning model;
Training unit is trained for input sample data and label into depth multicore mapping learning model;
Weighed value adjusting unit, for mapping the loss function of learning model according to the depth multicore after training to the depth
The weighted value of each basic core is adjusted in multicore mapping learning model, until loss function value is less than preset value.
Optionally, the model construction module, specifically includes:
Basic core determination unit, for determining the type and inner parameter of each basic core;
Model construction unit, for constructing using each basic core as the depth confidence network of the full connection structure of unit,
It is denoted as depth multicore mapping learning model.
Optionally, the weighed value adjusting unit, specifically includes:
Weighed value adjusting subelement is used for basisTo each in depth multicore mapping learning model
The weighted value of basic core is adjusted, whereinFor the weighted value of k-th in the t times iterative learning basic core, γkIt is k-th
The learning rate of basic core, TSpanTo pass through the loss function for staying a hair to obtain, θkFor the weight of k-th of basic core.
Optionally, the system also includes:
Test module, for determining the precision of the mapping learning model of the depth multicore after training using test data;
Precision judgment module, for judging whether the precision is less than given threshold;If it is not, then continuing to the depth
Multicore mapping learning model is trained.
The specific embodiment provided according to the present invention, the invention discloses following technical effects: image provided by the invention
The thought that feature extracting method and system are combined using multicore mapping and deep learning, will be in Remote Sensing Target feature extraction
Multicore mapping algorithm introduces deep structure, crosses the weight that depth network determines each basic core, and depth structure is traditional nuclear mapping
Method brings better Nonlinear feature extraction, and the feature obtained from remote sensing images is enabled preferably to describe target,
While effect and efficiency that lifting feature extracts, Characteristics of The Remote Sensing Images extraction field is also better achieved and has manually been set from dependence
Determine transformation of the mode of feature extraction target signature to the extraction mode for intelligently obtaining more accurate feature from data sample.
Detailed description of the invention
It in order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, below will be to institute in embodiment
Attached drawing to be used is needed to be briefly described, it should be apparent that, the accompanying drawings in the following description is only some implementations of the invention
Example, for those of ordinary skill in the art, without creative efforts, can also obtain according to these attached drawings
Obtain other attached drawings.
Fig. 1 is image characteristic extracting method of embodiment of the present invention flow diagram;
Fig. 2 is the network structure of depth of embodiment of the present invention nuclear mapping;
Fig. 3 is the basic depth confidence multicore mapping structure figure of the embodiment of the present invention;
Fig. 4 is ten class battlebus target of embodiment of the present invention imaging effect in SAR radar;
Fig. 5 is ten class battlebus target of the embodiment of the present invention in practical natural light effect;
Fig. 6 is image characteristic extraction system of embodiment of the present invention structural schematic diagram.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
The object of the present invention is to provide a kind of image characteristic extracting method and systems, have feature extraction accuracy height, effect
The high feature of rate.
In order to make the foregoing objectives, features and advantages of the present invention clearer and more comprehensible, with reference to the accompanying drawing and specific real
Applying mode, the present invention is described in further detail.
Multicore model is the stronger kernel-based learning algorithms model of a kind of flexibility, recent theory and application verified utilization
Multicore replaces monokaryon that can enhance the interpretation of decision function, and can obtain performance more preferably than monokaryon model.Multicore model is constructed,
A kind of most common method is exactly to consider the convex combination of multiple basic kernel functions, shaped like:
Here KjIt is basic kernel function, M is the total number of basic core, βjIt is weight coefficient.Therefore, under multicore frame, sample is in feature
Expression problem in space transforms into the select permeability of basic core and weight coefficient.In the group that this is constructed by multiple feature spaces
Close in space, due to combining the Feature Mapping ability of each basic core of being utilized, well solved kernel function selection and with
The selection problem of the relevant variable of core goal-griven metric and model.Meanwhile it is defeated by distinguishing the different characteristic component of isomeric data
Enter corresponding kernel function to be mapped, express that data preferably in new feature space, classification can significantly improved just
True rate or precision of prediction.For weight coefficient in multicore model determination invention introduces depth confidence network, realize pair
The optimization of multicore model.
Fig. 1 is image characteristic extracting method of embodiment of the present invention flow diagram, as shown in Figure 1, figure provided by the invention
As feature extracting method step is specific as follows:
Step 101: building depth multicore maps learning model, and each unit of learning model is a basic core, learns
Practising each basic core of model interlayer is full connection structure, and each basic core is mutually indepedent in layer;
Step 102: depth multicore mapping learning model being trained using sample data, sample data is the more of image
Dimensional feature data;
Step 103: obtaining images to be recognized;
Step 104: feature extraction is carried out to images to be recognized using the depth multicore mapping learning model after training.
Wherein, image characteristic extracting method provided by the invention further includes determining that the depth after training is more using test data
The precision of nuclear mapping learning model;Judge whether the precision is less than given threshold;If it is not, then continuing to the depth multicore
Mapping learning model is trained.
In addition, step 101 specifically includes:
Determine the type and inner parameter of each basic core;
Building is denoted as depth multicore mapping study mould using each basic core as the depth confidence network of the full connection structure of unit
Type.
Step 102 specifically includes:
The initial learning rate and initial weight value of set depth multicore mapping learning model;
Input sample data and label into depth multicore mapping learning model, are trained;
According to the loss function of the depth multicore mapping learning model after training to each in depth multicore mapping learning model
The weighted value of basic core is adjusted, until loss function value is less than preset value.
Detailed process is as follows:
Using the depth core learning structure for having hidden layer, formula expression are as follows:
k(xi,xj|θ)→k(g(xi,w),g(xj,w)|θ,w) (1)
Wherein, g (xj, w) and it is the nonlinear characteristic figure obtained by depth structure.
Fig. 2 is the network structure of depth of embodiment of the present invention nuclear mapping, firstly, single nuclear mapping is extended to linearly connected
Coenocytism, construct depth multicore mapping feature extraction structure, as shown in Figure 3.
In the depth multicore mapping structure of Fig. 3, using the full connection structure of depth confidence network, each unit is one
Basic core, each basic core in layer are independent from each other, and are full connection structures between each basic core of interlayer.Therefore, in training
It is also the weight coefficient according to the basic core of method optimizing of hierarchical optimization.Entire depth structure is functionally the equal of a core
The expansion of mapping is improved in feature extraction performance.The nesting that depth multicore mapping mode can be regarded as kernel function is come
Realize multilevel extension, l layers of nuclear mapping expression formula are as follows:
K(l)(x, y)=Φ(l)(...Φ(1)(x))·Φ(l)(...Φ(1)(y)) (2)
Wherein, Φ(1)It (x) is the 1st Nonlinear Mapping for sample x, Φ(l)(...Φ(1)It (y)) is the l layer of sample
Mapping.
When every layer of nuclear mapping is extended to the combination of multiple basic cores by monokaryon, that is, realize multilayer multicore mapping net
Network, expression formula are as follows:
Wherein,Indicate m-th of basic core in h-th of l row set,It is the corresponding weighted value of basic core, this
The basic depth coenocytism that invention uses is based on Fig. 3 and the deep structure that is adjusted to the number of plies and unit number, using damage
Lose the performance of function evaluation network parameter data sorting system adjusted.Basic depth coenocytism is similar in connection type
In depth confidence network, equally using connecting entirely between different layer units, structure connectionless between unit, difference exist in layer
Optimize network by optimizing the probability between each connection unit in, depth confidence network, and each unit of depth coenocytism
It is made of a basic core.Each basic core selectes its type and inner parameter before training, in the training process each iteration
Using the strategy of hierarchical optimization, the weight optimal solution of each basic core is obtained, right value update formula is as follows:
Wherein,For the weighted value of k-th in the t times iterative learning basic core, γkFor the learning rate of k-th of basic core,
TSpanTo pass through the loss function for staying a hair to obtain, θkFor the weight of k-th of basic core, loss function formula are as follows:
Data xpHigher dimensional space is mapped to by nuclear mapping mode to obtainSpIt is a littleTo set Γp
Distance, wherein
Wherein, λiIndicate feature combination coefficient,Indicate sample xiIn kernel function KθSample space under mapping.
Ten class battlebus targets are as data set under selection different angle, firstly, original image in MSTAR data set is carried out
Pretreatment, is cropped to consistent size, and then, choosing wherein 1000 pictures, as training set, 1000 pictures, which are used as, to be surveyed
Examination collection.In subsequent experimental, influence for test data set sample size to feature extraction result is constant in other conditions
In the case of, comparative test has been carried out using 400 training sets and 400 test sets as data object.Ten class battlebus targets are in SAR
Imaging effect and practical form such as Fig. 4-5 institute under the conditions of natural light in radar, Fig. 4,5 be be utilized it is provided by the invention
The image information that method proposes, from Fig. 4,5 as can be seen that can clearly tell battlebus mesh from the image that the present invention extracts
Mark.
The feature extraction effect under both of which is verified by construction different data collection, and by depth nuclear mapping feature extraction
Target classification result afterwards is compared with other common methods.Feature extraction effect is verified by classification task first, it is general
Logical SVM only supports two mode classifications, therefore, combination of two is carried out for different classes of battlebus target, then according to combination
Structure constructs data set, and depth nuclear mapping structure feature is verified under two different data set scales and extracts the effect in classification
Fruit, the results are shown in Table 1:
1 depth nuclear mapping target classification result of table
By five group of two classification task, the feature of depth nuclear mapping is mentioned for different objects and different size of data set
Result is taken to be verified.Same set of network parameters shows slightly difference, but whole accuracy rate in the classification task of different target
All as constructional depth is improved, and data set includes that increasing for data volume can improve classification accuracy, but calculate
Time can also increase accordingly, and experimental result is as shown in table 2:
2 depth nuclear mapping object detection results of table
Five classification target detection accuracy have differences, and illustrate that the otherness between target of all categories and other classifications is different,
There is also differences for the performance of the extracted feature of different target, but with the increase of constructional depth, detection performance is also obtained
It is promoted, illustrates that depth structure is able to ascend ability of the nuclear mapping in Characteristics of The Remote Sensing Images extraction.
Under same computing resource and data object, it is compared in mesh to instantly common target's feature-extraction mode
The classification performance in classification task is marked, what is selected in monokaryon mapping is using the support vector machines of RBF core, the choosing of depth confidence network
What is selected is the preferable four-layer network network structure of effect, and what convolutional network was chosen is common AlexNet model, each method classification results
It is as shown in table 3:
3 depth nuclear mapping classification results of table and other algorithms compare
By compare based on depth nuclear mapping extract feature classification results and other modes classification results it can be found that
Depth multicore mappings characteristics extraction effect is best.Accuracy rate has greatly improved compared with common single nuclear mapping algorithm, says
Bright depth structure can promote the feature extraction performance of nuclear mapping algorithm.
The thought that image characteristic extracting method provided by the invention is combined using multicore mapping and deep learning, by remote sensing figure
As the multicore mapping algorithm introducing deep structure in target's feature-extraction, the weight of each basic core is determined by depth network, it is deep
Degree structure is that traditional core mapping method brings better Nonlinear feature extraction, enables the feature obtained from remote sensing images
Characteristics of The Remote Sensing Images has also been better achieved while the effect and efficiency that lifting feature extracts in enough preferably description targets
Extraction field is from the mode by artificial setting feature extraction target signature to intelligently obtaining more accurate spy from data sample
The transformation of the extraction mode of sign.
The present invention also provides a kind of image characteristic extraction systems, as shown in fig. 6, the system includes:
Model construction module 601, for constructing depth multicore mapping learning model, each unit of learning model is one
A basis core, each basic core of learning model interlayer are full connection structure, and each basic core is mutually indepedent in layer;
Model training module 602, for being trained using sample data to depth multicore mapping learning model, sample number
According to the multi-dimensional feature data for image;
Image collection module 603, for obtaining images to be recognized;
Characteristic extracting module 604, for being carried out using the depth multicore mapping learning model after training to images to be recognized
Feature extraction.
Wherein, model construction module 601 specifically include:
Basic core determination unit, for determining the type and inner parameter of each basic core;
Model construction unit is denoted as constructing the depth confidence network using each basic core as the full connection structure of unit
Depth multicore maps learning model.
Model training module 602, specifically includes:
Initialization unit, initial learning rate and initial weight value for set depth multicore mapping learning model;
Training unit is trained for input sample data and label into depth multicore mapping learning model;
Weighed value adjusting unit, for mapping the loss function of learning model according to the depth multicore after training to depth multicore
The weighted value of each basic core is adjusted in mapping learning model, until loss function value is less than preset value.Weighed value adjusting unit,
Specifically include: weighed value adjusting subelement is used for basisTo each base in depth multicore mapping learning model
The weighted value of plinth core is adjusted, whereinFor the weighted value of k-th in the t times iterative learning basic core, γkFor k-th of base
The learning rate of plinth core, TSpanTo pass through the loss function for staying a hair to obtain, θkFor the weight of k-th of basic core.
Image characteristic extraction system provided by the invention further include:
Test module, for determining the precision of the mapping learning model of the depth multicore after training using test data;
Precision judgment module, for judging whether precision is less than given threshold;If it is not, then continuing to map depth multicore
Learning model is trained.
The thought that image characteristic extraction system provided by the invention is combined using multicore mapping and deep learning, by remote sensing figure
As the multicore mapping algorithm introducing deep structure in target's feature-extraction, the weight that depth network determines each basic core, depth are crossed
Structure is that traditional core mapping method brings better Nonlinear feature extraction, enables the feature obtained from remote sensing images
Preferably description target has also been better achieved Characteristics of The Remote Sensing Images and has mentioned while the effect and efficiency that lifting feature extracts
Take field from the mode by artificial setting feature extraction target signature to intelligently obtaining more accurate feature from data sample
Extraction mode transformation.
Each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with other
The difference of embodiment, the same or similar parts in each embodiment may refer to each other.For system disclosed in embodiment
For, since it is corresponded to the methods disclosed in the examples, so being described relatively simple, related place is said referring to method part
It is bright.
Specific examples are applied in the present invention, and principle and implementation of the present invention are described, above embodiments
Illustrate to be merely used to help understand method and its core concept of the invention;At the same time, for those skilled in the art, according to
According to thought of the invention, there will be changes in the specific implementation manner and application range.In conclusion the content of the present specification
It should not be construed as limiting the invention.
Claims (10)
1. a kind of image characteristic extracting method characterized by comprising
It constructs depth multicore and maps learning model, each unit of the learning model is a basic core, the study mould
Each basic core of type interlayer is full connection structure, and each basic core is mutually indepedent in layer;
Depth multicore mapping learning model is trained using sample data, the sample data is that the multidimensional of image is special
Levy data;
Obtain images to be recognized;
Feature extraction is carried out to the images to be recognized using the depth multicore mapping learning model after training.
2. image characteristic extracting method according to claim 1, which is characterized in that described to use sample data to the depth
Degree multicore mapping learning model is trained, and is specifically included:
Set the initial learning rate and initial weight value of the depth multicore mapping learning model;
Input sample data and label into depth multicore mapping learning model, are trained;
According to the loss function of the depth multicore mapping learning model after training to each in depth multicore mapping learning model
The weighted value of basic core is adjusted, until loss function value is less than preset value.
3. image characteristic extracting method according to claim 1, which is characterized in that the building depth multicore mapping study
Model specifically includes:
Determine the type and inner parameter of each basic core;
Building is denoted as depth multicore mapping study mould using each basic core as the depth confidence network of the full connection structure of unit
Type.
4. image characteristic extracting method according to claim 2, which is characterized in that the depth multicore according to after training
The loss function for mapping learning model is adjusted the weighted value of each basic core in depth multicore mapping learning model, has
Body includes:
According toThe weighted value of each basic core in depth multicore mapping learning model is adjusted,
Wherein,For the weighted value of k-th in the t times iterative learning basic core, γkFor the learning rate of k-th of basic core, TSpanIt is logical
Cross the loss function for staying a hair to obtain, θkFor the weight of k-th of basic core.
5. image characteristic extracting method according to claim 1, which is characterized in that the method also includes:
The precision of the depth multicore mapping learning model after training is determined using test data;
Judge whether the precision is less than given threshold;
If it is not, then continuing to be trained depth multicore mapping learning model.
6. a kind of image characteristic extraction system characterized by comprising
Model construction module, for constructing depth multicore mapping learning model, each unit of the learning model is one
Basic core, each basic core of the learning model interlayer are full connection structure, and each basic core is mutually indepedent in layer;
Model training module, for being trained using sample data to depth multicore mapping learning model, the sample
Data are the multi-dimensional feature data of image;
Image collection module, for obtaining images to be recognized;
Characteristic extracting module, for carrying out feature to the images to be recognized using the depth multicore mapping learning model after training
It extracts.
7. image characteristic extraction system according to claim 6, which is characterized in that the model training module, it is specific to wrap
It includes:
Initialization unit, for setting the initial learning rate and initial weight value of the depth multicore mapping learning model;
Training unit is trained for input sample data and label into depth multicore mapping learning model;
Weighed value adjusting unit, for mapping the loss function of learning model according to the depth multicore after training to the depth multicore
The weighted value of each basic core is adjusted in mapping learning model, until loss function value is less than preset value.
8. image characteristic extraction system according to claim 6, which is characterized in that the model construction module is specific to wrap
It includes:
Basic core determination unit, for determining the type and inner parameter of each basic core;
Model construction unit is denoted as constructing using each basic core as the depth confidence network of the full connection structure of unit
Depth multicore maps learning model.
9. image characteristic extraction system according to claim 7, which is characterized in that the weighed value adjusting unit, it is specific to wrap
It includes:
Weighed value adjusting subelement is used for basisTo each basis in depth multicore mapping learning model
The weighted value of core is adjusted, whereinFor the weighted value of k-th in the t times iterative learning basic core, γkFor k-th of basis
The learning rate of core, TSpanTo pass through the loss function for staying a hair to obtain, θkFor the weight of k-th of basic core.
10. image characteristic extraction system according to claim 7, which is characterized in that the system also includes:
Test module, for determining the precision of the mapping learning model of the depth multicore after training using test data;
Precision judgment module, for judging whether the precision is less than given threshold;If it is not, then continuing to the depth multicore
Mapping learning model is trained.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811525051.0A CN109635738A (en) | 2018-12-13 | 2018-12-13 | A kind of image characteristic extracting method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811525051.0A CN109635738A (en) | 2018-12-13 | 2018-12-13 | A kind of image characteristic extracting method and system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109635738A true CN109635738A (en) | 2019-04-16 |
Family
ID=66073565
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811525051.0A Pending CN109635738A (en) | 2018-12-13 | 2018-12-13 | A kind of image characteristic extracting method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109635738A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111639651A (en) * | 2019-12-26 | 2020-09-08 | 珠海大横琴科技发展有限公司 | Ship retrieval method and device based on full-connection layer feature extraction |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105488536A (en) * | 2015-12-10 | 2016-04-13 | 中国科学院合肥物质科学研究院 | Agricultural pest image recognition method based on multi-feature deep learning technology |
CN108416318A (en) * | 2018-03-22 | 2018-08-17 | 电子科技大学 | Diameter radar image target depth method of model identification based on data enhancing |
WO2018184208A1 (en) * | 2017-04-07 | 2018-10-11 | Intel Corporation | Methods and apparatus for deep learning network execution pipeline on multi-processor platform |
CN108764316A (en) * | 2018-05-18 | 2018-11-06 | 河海大学 | Remote sensing images scene classification method based on depth convolutional neural networks and Multiple Kernel Learning |
-
2018
- 2018-12-13 CN CN201811525051.0A patent/CN109635738A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105488536A (en) * | 2015-12-10 | 2016-04-13 | 中国科学院合肥物质科学研究院 | Agricultural pest image recognition method based on multi-feature deep learning technology |
WO2018184208A1 (en) * | 2017-04-07 | 2018-10-11 | Intel Corporation | Methods and apparatus for deep learning network execution pipeline on multi-processor platform |
CN108416318A (en) * | 2018-03-22 | 2018-08-17 | 电子科技大学 | Diameter radar image target depth method of model identification based on data enhancing |
CN108764316A (en) * | 2018-05-18 | 2018-11-06 | 河海大学 | Remote sensing images scene classification method based on depth convolutional neural networks and Multiple Kernel Learning |
Non-Patent Citations (3)
Title |
---|
何秀玲等: "融合特征基于深度多核学习的动态表情识别", 《计算机应用与软件》 * |
李玉鑑等: "基于多层感知器的深度核映射支持向量机", 《北京工业大学学报》 * |
杜京义: "基于核算法的故障智能诊断理论及方法研究", 《中国博士学位论文全文数据库 工程利技II辑》 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111639651A (en) * | 2019-12-26 | 2020-09-08 | 珠海大横琴科技发展有限公司 | Ship retrieval method and device based on full-connection layer feature extraction |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106156744B (en) | SAR target detection method based on CFAR detection and deep learning | |
CN106469299B (en) | A kind of vehicle search method and device | |
CN105512680B (en) | A kind of more view SAR image target recognition methods based on deep neural network | |
CN108038445B (en) | SAR automatic target identification method based on multi-view deep learning framework | |
CN105574505B (en) | The method and system that human body target identifies again between a kind of multiple-camera | |
CN105426914B (en) | A kind of image similarity detection method of facing position identification | |
CN106203483B (en) | A kind of zero sample image classification method based on semantic related multi-modal mapping method | |
CN106951915B (en) | One-dimensional range profile multi-classifier fusion recognition method based on category confidence | |
CN105718940B (en) | The zero sample image classification method based on factorial analysis between multiple groups | |
CN109785371A (en) | A kind of sun image method for registering based on normalized crosscorrelation and SIFT | |
CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
CN110084320A (en) | Thyroid papillary carcinoma Ultrasound Image Recognition Method, device, system and medium | |
CN103218825A (en) | Quick detection method of spatio-temporal interest points with invariable scale | |
CN109903339A (en) | A kind of video group personage's position finding and detection method based on multidimensional fusion feature | |
CN108830172A (en) | Aircraft remote sensing images detection method based on depth residual error network and SV coding | |
CN105205807A (en) | Remote sensing image change detection method based on sparse automatic code machine | |
CN109635738A (en) | A kind of image characteristic extracting method and system | |
CN114170526A (en) | Remote sensing image multi-scale target detection and identification method based on lightweight network | |
US6754390B2 (en) | Fusing outputs from multiple detection/classification schemes | |
CN113420593A (en) | Small sample SAR automatic target recognition method based on hybrid inference network | |
CN115100428A (en) | Target detection method using context sensing | |
Jeon et al. | Discovering latent topics with saliency-weighted LDA for image scene understanding | |
CN103279579B (en) | The video retrieval method in view-based access control model space | |
CN108985445A (en) | A kind of target bearing SAR discrimination method based on machine Learning Theory | |
CN110135280A (en) | A kind of multiple view SAR automatic target recognition method based on sparse representation classification |
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: 20190416 |
|
RJ01 | Rejection of invention patent application after publication |