CN107239759A - A kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic - Google Patents
A kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic Download PDFInfo
- Publication number
- CN107239759A CN107239759A CN201710387269.3A CN201710387269A CN107239759A CN 107239759 A CN107239759 A CN 107239759A CN 201710387269 A CN201710387269 A CN 201710387269A CN 107239759 A CN107239759 A CN 107239759A
- Authority
- CN
- China
- Prior art keywords
- image
- remote sensing
- depth characteristic
- source domain
- neural networks
- 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.)
- Granted
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
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
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 present invention provides a kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic.This method can directly be classified using existing remote sensing images and sample information to the remote sensing images newly obtained, so as to provide support for the fast slowdown monitoring of remote sensing.This method comprises the following steps:To source domain image and target area image, using principal component transform, its first three principal component component is extracted respectively;To the new images of three wave bands of generation, the image block centered on each pixel of extraction is input to the multilayer convolutional neural networks trained;Last full articulamentum of convolutional neural networks is exported, the depth characteristic for obtaining the pixel is represented;To source domain image and the training sample of source domain, the depth characteristic based on extraction obtains a grader using support vector machine classifier training;To target area image, directly classified using obtained supporting vector grader, completed from source domain image and classification corresponding relation to the transfer learning of aiming field.
Description
Technical field
The present invention relates to Remote Sensing Image Processing Technology, specifically, it is related to a kind of high-space resolution based on depth characteristic
Rate remote sensing images transfer learning method, this method can be right based on existing Hi-spatial resolution remote sensing image and sample data
The remote sensing images newly obtained are classified automatically in the case of no sample, are remote sensing prison so as to quickly handle remote sensing images
Survey and support is provided.
Background technology
Remote sensing technology is widely used in geoscience applications at present, and such as Forestry resource plan, Crop Estimation, environment are commented
Estimate, disaster monitoring etc..Remote Image Classification is the committed step for remote sensing images being converted to from data information.From whether needing
Want training sample to divide, Classifying Method in Remote Sensing Image is divided into supervised classification and unsupervised classification.It is most normal in remote sensing image classification
It is supervised classification method.Supervised classification method is, it is necessary to artificially determine taxonomic hierarchies, training sample of each classification etc..
Supervised classification sorting technique, which ensure that, obtains relatively good nicety of grading, but acquisition training sample is one and taken time and effort
Process.In normal mode, remote sensing satellite can be imaged to a certain given area according to certain cycle of returning to, and formed one and be
The remote sensing images of row time series.If the collection to all carrying out sample per the phase remote sensing images newly obtained, artificial behaviour
Work will be extremely tedious.Especially for the application such as remote sensing calamity emergency, it is necessary to quick in the case where obtaining remote sensing images
Ground obtains its thematic classification and change information.Thus, how research utilizes remote sensing images and the training sample letter obtained in the past
Breath, carrying out classification (resurveying for sample need not be carried out by referring to) automatically to the remote sensing images newly obtained has certain practical valency
Value.
The transfer learning in machine learning field is intended to satisfy that above-mentioned application demand.Transfer learning is appointed using some classification
Rule that business (including initial data and sample data of all categories, commonly referred to as source domain) study is obtained, rule to one newly
And with certain correlation classification task (only include initial data, without or seldom sample data, be also not enough to
Train effective a grader, commonly referred to as aiming field) classified.Current transfer learning algorithm mainly includes two major classes
Method:The method that the method and feature based of Case-based Reasoning are extracted.In former approach, its basic thought is, although source
The training data in domain and the data of aiming field are more or less different, but are due to that the data in two domains have certain correlation,
Thus the training data of source domain or should can have a part of data (such as one of data subset) and be adapted to mesh
Mark numeric field data trains an effective disaggregated model;Thus the target of such algorithm is exactly to find out those from source domain to be adapted to mesh
The example of numeric field data is marked, and these examples are classified for auxiliary mark domain, typical algorithm such as TrAdaBoost.
In two kinds of methods, its main thought is, by Feature Dimension Reduction, source domain data and target numeric field data, to divide and drop to maximum similar
On the subspace of property:On common subspace, the training sample of source domain is that can be directly used for the instruction that aiming field carries out grader
Practice, typical algorithm has TPLSA algorithms, CoCC algorithms etc..Although this two classes algorithm achieves certain effect in the application, its
Still there is a certain distance from practical application.
Top-level meeting 2012 Conference on Neurals of the Krizhevsky A. in 2012 in machine learning field
The AlexNet that Information Processing Systems (NIPS) are delivered is operated on the image recognition tasks of classics and taken
Important breakthrough is obtained, is improved than work discrimination of the tradition based on SVMs close to 10%, causes academia and industry
The great interest of boundary again to relatively classical neutral net.In academia, neutral net is one in 1980s
The machine learning school of individual main flow.But, due to lacking enough training datas and relatively weak computing capability at that time,
Training and its difficulty to multilayer neural network, the training time are long, and the often sorter network of generation over-fitting, thus
Recognition effect in practical application is simultaneously bad.In recent years, image data set with enough big data quantities, with category label
Occur (such as ImageNet), and video card GPU computing capabilitys enhancing, enabling to multilayer in the acceptable time
The network model parameter that neural metwork training has gone out.The most basic construction unit of convolutional neural networks, comprising:Convolutional layer, Chi Hua
Layer and full articulamentum.Convolutional layer carries out convolution operation to the image in window, extracts various types of features;Pond layer typical case
Operation include average and maximize, be to extracting the further abstract of feature;Full articulamentum is by convolutional layer and pond layer
Output is stacked up using vector, is formed one or more layers full articulamentum, is realized the derivation ability of high-order.AlexNet obtains huge
After ten-strike, follow-up study person is again widely studied in each side such as the increase network number of plies, network optimized approach, occurs in that again
GoogleNet, VGGNet etc. influential convolutional neural networks.Researcher and designed network structure, and
The network model parameter trained is disclosed, and is directly used for follow-up researcher, or for the data of oneself
Collection carries out the adjusting and optimizing of parameter.
It is pointed out that these disclosed classical convolutional neural networks (AlexNet, GoogleNet, VGGNet etc.) are all
It is to be obtained by being trained to natural target image (the various natural familiar objects of such as automobile, aircraft and various scenes)
, thus often can not directly be applied in remote sensing images.Hi-spatial resolution remote sensing image has with these natural scenes
Certain similitude, but more is different;Other identification mission is also differed, such as Hi-spatial resolution remote sensing image is main
For distinguishing the ground mulching classifications such as building, road, trees, meadow.The basic ideas of this patent are, the remote sensing figure based on history
As (remotely-sensed data of such as previous phase) and sample data, using multilayer convolutional neural networks (such as AlexNet,
GoogleNet, VGGNet etc.) its last full articulamentum feature is extracted, it is used as its depth characteristic;And utilize sample data base
The training of grader is carried out in the depth characteristic of extraction;To the remote sensing images newly obtained, using same convolutional neural networks, carry
Its depth characteristic is taken, and is directly divided using the sorter model obtained based on history remote sensing images and sample data training
Class.
This patent proposes a kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic:For distant
The demand of sensing fast slowdown monitoring in using, based on existing Hi-spatial resolution remote sensing image and sample data, this patent is carried
The remote sensing images transfer learning method gone out, its core is to utilize the convolutional neural networks for widely using and getting immense success at present
Characteristics of The Remote Sensing Images is extracted, this characteristics of image belongs to the higher level spy of image due to being extracted by the neutral net of multilayer
The abstract expression levied, and it is not limited to the expression of traditional image spectrum feature, it is thus possible to reduce different phase remote sensing images
Radiation otherness to the influence that causes of classification, with stronger transfer ability.The technical method can utilize existing remote sensing
Image and sample information, train grader;To the remote sensing images newly obtained, directly entered using the grader trained
Row classification, so as to provide support for the application such as fast slowdown monitoring of remote sensing.
The content of the invention
It is an object of the invention to provide a kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic.
The present invention basic ideas be:For source domain image and target area image respectively using the convolution god trained
The extraction of depth characteristic is carried out through network;The depth characteristic based on source domain image zooming-out is divided using the training sample of source domain
The training of class device;The grader obtained to the training sample training based on source domain, the depth characteristic to aiming field image zooming-out is entered
Row Direct Classification, completes the migration of classificating knowledge.
A kind of Hi-spatial resolution remote sensing image transfer learning based on depth characteristic that technical scheme is provided
Method, it is characterised in that including following implementation steps:
The convolutional neural networks model that A selections have been trained;
B is to source domain image and target area image, using principal component transform, and its first three principal component component is extracted respectively,
Form the new images of corresponding three wave bands;
C takes the image of its certain size windows centered on each pixel respectively to the new images of three wave bands of generation
Block is input to selected convolutional neural networks;
Last full articulamentum in D output convolutional neural networks, the depth characteristic for obtaining the pixel is represented;
E uses SVMs to source domain image and the training sample of source domain based on the depth characteristic extracted in step D
Grader, which is trained, obtains a grader;
F is carried out to the depth characteristic of the target area image obtained in step D using obtained grader is trained in step E
Classification, is completed from source domain image and classification corresponding relation to the transfer learning of aiming field.
Above-mentioned implementation steps are characterised by:
The convolutional neural networks model trained in step A refers to AlexNet, VGGNet, GoogleNet etc. by using
The multilayer convolution that field of image recognition maximum, the database ImageNet training with a large amount of flag datas are obtained in the world at present
Neutral net.
Principal component transform described in step B be due to high spatial resolution remote sensing images be owned by greatly more than 3 into
As wave band (in addition to conventional three spectral bands of RGB, typically all there is the near infrared band of observation vegetation), such as IKONOS
Satellite has 4 spectral bands;And newest WorldView-4 satellites have 8 spectral bands;Need to use principal component analysis
Conventional Hi-spatial resolution remote sensing image is subjected to dimension-reduction treatment, its first three principal component component, dimensionality reduction to three ripples is obtained
Section.
A certain size window described in step C, chooses the window such as 5 × 5,7 × 7,9 × 9,11 × 11 of odd sized
Deng.Window size is not answered excessive, it is necessary to depending on according to the spatial resolution and application demand of remote sensing images;General selection is former
It is then set window size, makes pixel in the image block that it chooses should be largely (more than 80%) and center pixel
Type of ground objects is consistent.
The depth characteristic of the pixel described in step D is represented, refers to choose each pixel the image centered on it
Block, is input to after convolutional neural networks, is gradually calculated backward since the first layer network, and to the last a full articulamentum, is obtained
To the vector of a higher-dimension, the depth characteristic of the pixel is used as.
Step E refers to that the depth characteristic and sample class information of the higher-dimension obtained using source domain image use classics
Support vector machine classifier is trained, and obtains a grader;SVMs Selection of kernel function linear kernel function.
Step F refers to, using training obtained grader, the depth for the higher-dimension that input target area image is obtained in step E
Feature, obtains the classification results of target area image, completes the migration from source domain image knowledge to target area image.
The present invention has following features compared with prior art:The Hi-spatial resolution remote sensing image transfer learning algorithm is abundant
The convolutional neural networks of the multilayer of current comparative maturity are make use of, so that the high-level feature of image is captured, so as to avoid
The otherness of the image light spectrum that causes of the two width different images due to radiating difference, and then can preferably lift image migration
Results of learning, provide technical support, and then be the fast slowdown monitoring service of remote sensing for the new rapid automatized processing for obtaining image.
Brief description of the drawings:
Accompanying drawing is a kind of Hi-spatial resolution remote sensing image transfer learning method flow diagram based on depth characteristic
Embodiment:
A kind of reality of the Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic is realized using the present invention
Apply for example shown in accompanying drawing, be described in conjunction with accompanying drawing.
The current common multiwave Hi-spatial resolution remote sensing image (IKONOS of such as four wave bands of 100 pairs of processing unit
Satellite multispectral image, the WorldView-4 satellites multispectral image of eight wave bands) principal component transform is carried out, use first three
Component (first three component for most having information content) generates the image of three wave band.
Processing unit 101 reads odd number using it as geometric center to each pixel of the image obtained in processing unit 100
(such as 5 × 5,7 × 7,9 × 9,11 × image block 11), the image block has spatial context letter to window size as the pixel
The expression of breath.
Processing unit 102 inputs existing convolutional neural networks, i.e., image block processing unit 101 extracted is input to
Trained good convolutional neural networks (such as AlexNet, GoogleNet, VGGNet etc.).Convolutional neural networks typically have two species
The operation of type:Propagated forward is operated and backpropagation operation.Reverse operating is substantially carried out the transmission backward of derivative, for network
The parameter learning of structure.Due to we make use of the convolutional neural networks trained (such as AlexNet, GoogleNet,
VGGNet etc.), we need not carry out reverse operation, i.e., need not carry out the study of network parameter.
We are only needed to before convolutional neural networks are carried out to operation.During propagated forward, we are image
Block is input in convolutional neural networks, is successively calculated, until calculating to last full articulamentum.
It is complete that processing unit 103 extracts last of the characteristic vector of one group of higher-dimension i.e. the convolutional neural networks used
The result output of articulamentum, the vectorial dimension determines that (such as AlexNet is most by the structure of the convolutional neural networks used
Latter full articulamentum, i.e. FC7 layer, 4096) vector dimension is.To each pixel, the image extracted by processing unit 101
Block is input to convolutional neural networks, can all export the characteristic vector of a higher-dimension, and the vector is high-level abstract of the pixel
Character representation, is its depth characteristic.
The Training Support Vector Machines grader of processing unit 104 is i.e. according to existing taxonomic hierarchies, sample data of all categories
And the picture depth feature of the extraction of processing unit 103 carries out the training of grader, the classical SVMs of grader selection.
And vector of the feature of the extraction of processing unit 103 for higher-dimension is considered, important Selection of kernel function in support vector machine classifier
The problem of, from linear kernel function.Processing unit 104 is only handled the initial data and sample data of source domain.
The high spatial resolution remote sense figure that processing unit 105 is classified i.e. to newly obtaining using existing sorter model
Picture, the high dimensional feature vector extracted after processing unit 103, the SVMs point directly generated using processing unit 104
Class device, is predicted, the classification results of the remote sensing images newly obtained.The process need not be entered to the remote sensing images newly obtained
Row sample collection, so as to accelerate the flow of image classification.The key of the image transfer learning strategy, is to all remote sensing figures
As having used the convolutional neural networks that there is certain depth with many levels to carry out feature extraction, so that the image extracted
Feature has very high type of ground objects abstract expression ability, and then reduces the two images spectral value difference that radiation difference is caused
Influence to classification, so as in the case where not re-starting sample collection, to the image newly obtained, obtains preferable
Classifying quality.
An example of the present invention realizes on a pc platform, experiments verify that, the image transfer learning algorithm utilizes multilayer
Convolutional neural networks extract characteristics of image, existing Hi-spatial resolution remote sensing image and training sample are classified
The training of device, can obtain reliable sorter model, and the remote sensing images newly obtained can automatically be classified and (be not required to
Obtain extra training sample), classification results are reliable, can be used in the application demand of the fast slowdown monitoring of remote sensing.
It should be pointed out that embodiment described above can make those skilled in the art that this hair is more fully understood
It is bright, but do not limit the invention in any way.Therefore, it will be appreciated by those skilled in the art that still can be to present invention progress
Modification or equivalent substitution;And technical scheme and its improvement of all spirit and technical spirit that do not depart from the present invention, it all should
Cover among the protection domain of patent of the present invention.
Claims (6)
1. a kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic, it is characterised in that including following step
Suddenly:
The convolutional neural networks model that A selections have been trained;
B is to source domain image and target area image, using principal component transform, and its first three principal component component is extracted respectively, is formed
The new images of corresponding three wave bands;
C takes the image block of its certain size windows defeated the new images of three wave bands of generation centered on each pixel respectively
Enter to selected convolutional neural networks;
Last full articulamentum in D output convolutional neural networks, the depth characteristic for obtaining the pixel is represented;
E uses support vector cassification to source domain image and the training sample of source domain based on the depth characteristic extracted in step D
Device, which is trained, obtains a grader;
F is to the depth characteristic of the target area image obtained in step D, using training obtained grader to be classified in step E,
Complete from source domain image and classification corresponding relation to the transfer learning of aiming field.
2. according to the method described in claim 1, it is characterised in that be based on principal component transform in step B, for extracting its first three
Individual principal component component.
3. according to the method described in claim 1, it is characterised in that a certain size the window described in step C, choose odd number
The window of size;Meanwhile, window is chosen should not be excessive;Selection principle is set window size, the image block for choosing it
Interior pixel should be largely consistent with the type of ground objects of center pixel (more than 80%).
4. according to the method described in claim 1, it is characterised in that the depth characteristic of the pixel described in step D is represented, is
Finger is input to the image block in step C after convolutional neural networks, the output vector of the full articulamentum of last obtained.
5. according to the method described in claim 1, it is characterised in that to the depth characteristic and instruction of source domain image zooming-out in step E
Practice sample and carry out classifier training from SVMs.
6. according to the method described in claim 1, it is characterised in that the SVMs described in step E, the choosing of its kernel function
Select, from linear kernel function.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710387269.3A CN107239759B (en) | 2017-05-27 | 2017-05-27 | High-spatial-resolution remote sensing image transfer learning method based on depth features |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710387269.3A CN107239759B (en) | 2017-05-27 | 2017-05-27 | High-spatial-resolution remote sensing image transfer learning method based on depth features |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107239759A true CN107239759A (en) | 2017-10-10 |
CN107239759B CN107239759B (en) | 2020-11-27 |
Family
ID=59985239
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710387269.3A Active CN107239759B (en) | 2017-05-27 | 2017-05-27 | High-spatial-resolution remote sensing image transfer learning method based on depth features |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107239759B (en) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107909101A (en) * | 2017-11-10 | 2018-04-13 | 清华大学 | Semi-supervised transfer learning character identifying method and system based on convolutional neural networks |
CN108629289A (en) * | 2018-04-11 | 2018-10-09 | 千寻位置网络有限公司 | The recognition methods in farmland and system, applied to the unmanned plane of agricultural |
CN109003239A (en) * | 2018-07-04 | 2018-12-14 | 华南理工大学 | A kind of multispectral image sharpening method based on transfer learning neural network |
CN109359623A (en) * | 2018-11-13 | 2019-02-19 | 西北工业大学 | High spectrum image based on depth Joint Distribution adaptation network migrates classification method |
CN109740495A (en) * | 2018-12-28 | 2019-05-10 | 成都思晗科技股份有限公司 | Outdoor weather image classification method based on transfer learning technology |
CN110111297A (en) * | 2019-03-15 | 2019-08-09 | 浙江大学 | A kind of injection-molded item surface image defect identification method based on transfer learning |
CN110837875A (en) * | 2019-11-18 | 2020-02-25 | 国家基础地理信息中心 | Method and device for judging quality abnormity of earth surface coverage data |
CN111191510A (en) * | 2019-11-29 | 2020-05-22 | 杭州电子科技大学 | Relation network-based remote sensing image small sample target identification method in complex scene |
CN111274905A (en) * | 2020-01-16 | 2020-06-12 | 井冈山大学 | AlexNet and SVM combined satellite remote sensing image land use change detection method |
CN112232249A (en) * | 2020-10-22 | 2021-01-15 | 中国科学院空天信息创新研究院 | Remote sensing image change detection method and device based on depth features |
Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101794396A (en) * | 2010-03-25 | 2010-08-04 | 西安电子科技大学 | System and method for recognizing remote sensing image target based on migration network learning |
CN101976361A (en) * | 2010-11-23 | 2011-02-16 | 中国矿业大学 | Multi-kernel support vector machine classification method for remote sensing images |
CN102915445A (en) * | 2012-09-17 | 2013-02-06 | 杭州电子科技大学 | Method for classifying hyperspectral remote sensing images of improved neural network |
US20130132311A1 (en) * | 2011-11-18 | 2013-05-23 | Honeywell International Inc. | Score fusion and training data recycling for video classification |
US20130322742A1 (en) * | 2012-05-29 | 2013-12-05 | The Johns Hopkins University | Tactical Object Finder |
CN105469109A (en) * | 2015-11-19 | 2016-04-06 | 中国地质大学(武汉) | Transfer learning method based on class centroid alignment and for remote sensing image classification |
CN106023065A (en) * | 2016-05-13 | 2016-10-12 | 中国矿业大学 | Tensor hyperspectral image spectrum-space dimensionality reduction method based on deep convolutional neural network |
CN106096660A (en) * | 2016-06-22 | 2016-11-09 | 东南大学 | Convolutional neural networks based on independent composition analysis algorithm |
CN106326931A (en) * | 2016-08-25 | 2017-01-11 | 南京信息工程大学 | Mammary gland molybdenum target image automatic classification method based on deep learning |
CN106599922A (en) * | 2016-12-16 | 2017-04-26 | 中国科学院计算技术研究所 | Transfer learning method and transfer learning system for large-scale data calibration |
CN106650781A (en) * | 2016-10-21 | 2017-05-10 | 广东工业大学 | Convolutional neural network image recognition method and device |
-
2017
- 2017-05-27 CN CN201710387269.3A patent/CN107239759B/en active Active
Patent Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101794396A (en) * | 2010-03-25 | 2010-08-04 | 西安电子科技大学 | System and method for recognizing remote sensing image target based on migration network learning |
CN101976361A (en) * | 2010-11-23 | 2011-02-16 | 中国矿业大学 | Multi-kernel support vector machine classification method for remote sensing images |
US20130132311A1 (en) * | 2011-11-18 | 2013-05-23 | Honeywell International Inc. | Score fusion and training data recycling for video classification |
US20130322742A1 (en) * | 2012-05-29 | 2013-12-05 | The Johns Hopkins University | Tactical Object Finder |
CN102915445A (en) * | 2012-09-17 | 2013-02-06 | 杭州电子科技大学 | Method for classifying hyperspectral remote sensing images of improved neural network |
CN105469109A (en) * | 2015-11-19 | 2016-04-06 | 中国地质大学(武汉) | Transfer learning method based on class centroid alignment and for remote sensing image classification |
CN106023065A (en) * | 2016-05-13 | 2016-10-12 | 中国矿业大学 | Tensor hyperspectral image spectrum-space dimensionality reduction method based on deep convolutional neural network |
CN106096660A (en) * | 2016-06-22 | 2016-11-09 | 东南大学 | Convolutional neural networks based on independent composition analysis algorithm |
CN106326931A (en) * | 2016-08-25 | 2017-01-11 | 南京信息工程大学 | Mammary gland molybdenum target image automatic classification method based on deep learning |
CN106650781A (en) * | 2016-10-21 | 2017-05-10 | 广东工业大学 | Convolutional neural network image recognition method and device |
CN106599922A (en) * | 2016-12-16 | 2017-04-26 | 中国科学院计算技术研究所 | Transfer learning method and transfer learning system for large-scale data calibration |
Non-Patent Citations (2)
Title |
---|
FAN HU ET AL: "Transferring Deep Convolutional Neural Networks for the Scene Classification of High-Resolution Remote Sensing Imagery", 《REMOTE SENS》 * |
赵理君 等: "典型遥感数据分类方法的适用性分析——以遥感图像场景分类为例", 《遥感学报》 * |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107909101A (en) * | 2017-11-10 | 2018-04-13 | 清华大学 | Semi-supervised transfer learning character identifying method and system based on convolutional neural networks |
CN107909101B (en) * | 2017-11-10 | 2019-07-12 | 清华大学 | Semi-supervised transfer learning character identifying method and system based on convolutional neural networks |
CN108629289A (en) * | 2018-04-11 | 2018-10-09 | 千寻位置网络有限公司 | The recognition methods in farmland and system, applied to the unmanned plane of agricultural |
CN108629289B (en) * | 2018-04-11 | 2021-07-30 | 千寻位置网络有限公司 | Farmland identification method and system and agricultural unmanned aerial vehicle |
CN109003239A (en) * | 2018-07-04 | 2018-12-14 | 华南理工大学 | A kind of multispectral image sharpening method based on transfer learning neural network |
CN109003239B (en) * | 2018-07-04 | 2022-03-29 | 华南理工大学 | Multispectral image sharpening method based on transfer learning neural network |
CN109359623B (en) * | 2018-11-13 | 2021-05-11 | 西北工业大学 | Hyperspectral image migration classification method based on depth joint distribution adaptive network |
CN109359623A (en) * | 2018-11-13 | 2019-02-19 | 西北工业大学 | High spectrum image based on depth Joint Distribution adaptation network migrates classification method |
CN109740495A (en) * | 2018-12-28 | 2019-05-10 | 成都思晗科技股份有限公司 | Outdoor weather image classification method based on transfer learning technology |
CN110111297A (en) * | 2019-03-15 | 2019-08-09 | 浙江大学 | A kind of injection-molded item surface image defect identification method based on transfer learning |
CN110837875A (en) * | 2019-11-18 | 2020-02-25 | 国家基础地理信息中心 | Method and device for judging quality abnormity of earth surface coverage data |
CN111191510A (en) * | 2019-11-29 | 2020-05-22 | 杭州电子科技大学 | Relation network-based remote sensing image small sample target identification method in complex scene |
CN111191510B (en) * | 2019-11-29 | 2022-12-09 | 杭州电子科技大学 | Relation network-based remote sensing image small sample target identification method in complex scene |
CN111274905A (en) * | 2020-01-16 | 2020-06-12 | 井冈山大学 | AlexNet and SVM combined satellite remote sensing image land use change detection method |
CN112232249A (en) * | 2020-10-22 | 2021-01-15 | 中国科学院空天信息创新研究院 | Remote sensing image change detection method and device based on depth features |
CN112232249B (en) * | 2020-10-22 | 2023-08-15 | 中国科学院空天信息创新研究院 | Remote sensing image change detection method and device based on depth characteristics |
Also Published As
Publication number | Publication date |
---|---|
CN107239759B (en) | 2020-11-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107239759A (en) | A kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic | |
Khan et al. | An automated system for cucumber leaf diseased spot detection and classification using improved saliency method and deep features selection | |
Lee et al. | Contextual deep CNN based hyperspectral classification | |
CN107016405B (en) | A kind of pest image classification method based on classification prediction convolutional neural networks | |
CN106845418A (en) | A kind of hyperspectral image classification method based on deep learning | |
CN112801146B (en) | Target detection method and system | |
CN108090447A (en) | Hyperspectral image classification method and device under double branch's deep structures | |
CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
CN110458201B (en) | Object-oriented classification method and classification device for remote sensing image | |
Chowdhury et al. | A novel texture feature based multiple classifier technique for roadside vegetation classification | |
CN107145889A (en) | Target identification method based on double CNN networks with RoI ponds | |
CN113705641B (en) | Hyperspectral image classification method based on rich context network | |
Cai et al. | Residual-capsule networks with threshold convolution for segmentation of wheat plantation rows in UAV images | |
CN113609889B (en) | High-resolution remote sensing image vegetation extraction method based on sensitive characteristic focusing perception | |
Li et al. | A spectral-spatial kernel-based method for hyperspectral imagery classification | |
Faria et al. | Fusion of time series representations for plant recognition in phenology studies | |
CN110807485B (en) | Method for fusing two-classification semantic segmentation maps into multi-classification semantic map based on high-resolution remote sensing image | |
CN110111346B (en) | Remote sensing image semantic segmentation method based on parallax information | |
Hassam et al. | A single stream modified mobilenet v2 and whale controlled entropy based optimization framework for citrus fruit diseases recognition | |
Septiarini et al. | Image processing techniques for tomato segmentation applying k-means clustering and edge detection approach | |
CN117372853A (en) | Underwater target detection algorithm based on image enhancement and attention mechanism | |
CN116740516A (en) | Target detection method and system based on multi-scale fusion feature extraction | |
Khudov et al. | The method for determining informative zones on images from on-board surveillance systems | |
Hu et al. | Learning salient features for flower classification using convolutional neural network | |
Varkentin et al. | Development of an application for car license plates recognition using neural network technologies |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |