CN109711470A - Merge spectrum-spatial information (si) DBM hyperspectral image classification method - Google Patents
Merge spectrum-spatial information (si) DBM hyperspectral image classification method Download PDFInfo
- Publication number
- CN109711470A CN109711470A CN201811616702.7A CN201811616702A CN109711470A CN 109711470 A CN109711470 A CN 109711470A CN 201811616702 A CN201811616702 A CN 201811616702A CN 109711470 A CN109711470 A CN 109711470A
- Authority
- CN
- China
- Prior art keywords
- spectrum
- pixel
- dbm
- vector
- spatial
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
A kind of fusion spectrum-spatial information (si) depth Boltzmann machine DBM hyperspectral image classification method, include the following steps: S100: to spatial information (si) of its space neighborhood information of each image element extraction as the pixel in original high spectrum image, and combining the spectrum-spatial information (si) to form the pixel with its spectrum domain information;S200: feature extraction is carried out by spectrum-spatial information (si) of the depth Boltzmann machine DBM to the pixel;S300: classified using logistic regression LR to extracted feature, obtain the category label of each pixel.DBM is applied in classification hyperspectral imagery by this method, obtains stronger classification capacity, effectively improves high-spectrum remote sensing data nicety of grading.
Description
Technical field
The disclosure belongs to technical field of remote sensing image processing, in particular to a kind of fusion spectrum-spatial information (si) DBM EO-1 hyperion
Image classification method.
Background technique
With being constantly progressive for remotely sensed image technology, the spectral resolution of acquired remotely-sensed data has had reached nanometer
Rank.In past 20 years, high-spectrum remote sensing has become the valuable source in people's production and living, has been applied to and has been permitted
It is multi-field, such as crops monitoring, mineral composition detection, target detection etc..In these application fields, classification is a pass
Key technology, it, which refers to, assigns pixel each in high-spectrum remote sensing to a classification mark according to the atural object classification representated by it
Note.
In order to realize the classification of high-spectrum remote sensing, researcher proposes a large amount of classification method in succession, as K is nearest
Adjacent method, maximum likelihood method, neural network, support vector machine, random forest.These methods are shown in specific remote sensing images
Excellent classification capacity, but for atural object metamorphosis multiplicity in high-spectrum remote sensing, spectral band is more, dimension is high, with
And there are the characteristics that bulk redundancy, traditional mode recognition methods classification capacity is limited.
In recent years, deep learning method achieves the achievement attracted attention in numerous applications, as a kind of machine learning algorithm,
It for basic unit, passes through stacked system structure with limited Boltzmann machine (Restricted Boltzmann Machine, RBM)
Make a profound structure.RBM model can carry out Nonlinear Mapping to input data, therefore deep learning method can be learned
Practise the nonlinear dependence between data, the especially abstract representation of the different levels between data.Common deep learning
Model includes: autocoder (Auto encoder, AE), deepness belief network (Deep Belief Network, DBN), depth
Spend Boltzmann machine (Deep Boltzmann Machine, DBM) and convolutional neural networks etc..
In terms of classification hyperspectral imagery, it is applied there are many deep learning method.EO-1 hyperion based on AE model
Image scene classification method, although can learn the stack up deep learning model of building of multiple AE models to EO-1 hyperion number
According to depth characteristic, but its reconstruction information for laying particular emphasis on learning data, which has limited its discriminating powers in classification task;
Hyperspectral data classification method based on DBN model by unsupervised pre-training and has the trim process of supervision, a DBN model
It can learn the discriminant information to data.But it is come from since the node of each hidden layer in the pre-training algorithm of DBN model only receives
Below the information of node layer and have ignored the information of node layer above, to limit the energy of dependence between model learning data
Power;CNN model carries out the target detection of SAR image, which has translation invariance, and is able to solve lacking for partial target
Mistake problem.But for the high-spectral data for lacking marked sample, this method is easily trapped into over-fitting.
Spatial information (si) based on atural object is for a hot issue in remote sensing image classification and Research on Remote Sensing Classification.
Due to the space attribute that atural object itself has, the precision of classification certainly will be can be improved by merging spatial information (si) in spectral information.?
It is direct and convenient that spectral information is obtained in remote sensing images, and obtains spatial information (si) and then need special method, such as shape
State method, MRF method and neighborhood method etc..
In view of high-spectral data intrinsic high-dimensional property and redundancy the characteristics of, how to improve dividing for high-spectral data
Class precision is most important.
Summary of the invention
To solve the above-mentioned problems, present disclose provides a kind of fusion spectrum-spatial information (si) depth Boltzmann machine DBM high
Spectrum picture classification method, includes the following steps:
S100: believe in the airspace to its space neighborhood information of each image element extraction in original high spectrum image as the pixel
Breath, and combine the spectrum-spatial information (si) to form the pixel with its spectrum domain information;
S200: feature extraction is carried out by spectrum-spatial information (si) of the depth Boltzmann machine DBM to the pixel;
S300: classified using logistic regression LR to extracted feature, obtain the category label of each pixel.
Through the above technical solutions, firstly, this method merges it in the spectral information of each pixel of high spectrum image
Spatial information (si) can effectively improve nicety of grading.Secondly, this method carries out the feature extraction of high spectrum image using DBM model, by
Each layer of hidden node receives the information from adjacent two sides hidden node simultaneously in the model, thus in identical quantity hidden node
Under conditions of, the feature learnt has more identification, is conducive to later period classification.As it can be seen that for high spectrum image, this method
It is compared with other deep learning methods, there is stronger classification capacity.
Detailed description of the invention
Fig. 1 is a kind of depth Boltzmann machine DBM of fusion spectrum-spatial information (si) provided in an embodiment of the present disclosure
The flow diagram of hyperspectral image classification method;
Fig. 2 is the spatial feature vector that pixel in high spectrum image is constructed in an embodiment of the present disclosure;
Fig. 3 is the relationship of different hidden layer numbers and classification accuracy rate in DBM in an embodiment of the present disclosure;
Fig. 4 is the relationship of the number of hidden nodes and classification accuracy rate in DBM in an embodiment of the present disclosure;
Fig. 5 (a) is DBN model structure in an embodiment of the present disclosure;
Fig. 5 (b) is DBM model structure in an embodiment of the present disclosure;
Fig. 6 is in an embodiment of the present disclosure with the mistake point rate comparison diagram of the increase DBM and DBN of the number of iterations.
Specific embodiment
In one embodiment, as shown in Figure 1, disclosing a kind of fusion spectrum-spatial information (si) depth Boltzmann machine DBM
Hyperspectral image classification method includes the following steps:
S100: believe in the airspace to its space neighborhood information of each image element extraction in original high spectrum image as the pixel
Breath, and combine the spectrum-spatial information (si) to form the pixel with its spectrum domain information;
S200: feature extraction is carried out by spectrum-spatial information (si) of the depth Boltzmann machine DBM to the pixel;
S300: classified using logistic regression LR to extracted feature, obtain the category label of each pixel.
For the embodiment, it is direct that the spectrum domain information of pixel is extracted in high-spectrum remote sensing data.And atural object
Therefore the structural attribute that target has its intrinsic extracts the spatial information (si) of atural object for classifying in high-spectral data, can have
Effect improves the precision of classification.Then spectrum-the spatial information (si) for constructing pixel in conjunction with the spectrum domain information of pixel first passes through DBM
Feature extraction is carried out to comprehensive spectrum-spatial information (si), is finally classified with LR.As it can be seen that this method has stronger classification energy
Power can effectively improve high-spectrum remote sensing data nicety of grading.
In another embodiment, step S100 further comprises:
S101: the spatial feature vector h of pixel is constructedspa;
S102: the spectral signature vector h of pixel is extracted in original high spectrum imagespe;
S103: the spectral signature vector h of the pixelspeWith spatial feature vector hspaIt is special to merge into comprehensive spectrum-airspace
Levy vector h.
For the embodiment, construction spectrum-spatial feature vector is conducive to the specific implementation of subsequent classification method.
In another embodiment, as shown in Fig. 2, step 101 further comprises:
S1001: PCA whitening processing is carried out along spectrum direction to the original high spectrum image;
S1002: intercepting the neighborhood of c × c size of each pixel, and be drawn into one-dimensional vector, and obtaining length is c × c × k
Spatial feature vector hspa。
Where it is assumed that original hyperspectral image data is m × n × l three-dimensional matrice, wherein m, n are exactly high spectrum image
Length and width, l be the length of the spectral vector of pixel, the i.e. number of spectrum.Each pixel corresponds to the spectral signature that length is l
Vector hspe.The neighborhood element of c × c size of each pixel is taken, it corresponds to the three-dimensional local matrix of a c × c × l, by it
It is drawn into one-dimensional vector, to obtain the spatial feature vector h of pixelspa, the length is c × c × l.To high-spectrum remote sensing data
PCA whitening processing is carried out along spectrum direction, k principal component before retaining, thus matrix size boil down to m × n × k, wherein k < <
l.C is exactly the Size of Neighborhood of pixel.K is exactly the number of PCA principal component.The value range of k is between [1, l].The value range of c
The length and wide maximum value for arriving each pixel for 3.
For the embodiment, due to the higher-dimension and redundancy of hyperspectral image data, spatial feature vector dimension
It is excessively high.Therefore, PCA whitening processing first is carried out to high-spectral data before constructing spatial feature vector, reaches dimensionality reduction and de-redundant
Remaining purpose.
In another embodiment, the spectrum of construction-spatial feature vector h is input in DBM, carries out feature extraction.
In another embodiment, refer to can as DBM's using spectrum-spatial feature vector h for the process of the feature extraction
See a layer vector, inside DBM, obtained hidden layer vector is exactly extracted feature vector after upward reasoning.
For the embodiment, this feature vector is learnt by DBM, an integrated spectral and space can be obtained
The expression of structure feature.The advantage is that: (1) can simultaneously compress spectral domain and spatial information (si), image element information is obtained
Effectively indicate;(2) feature extraction is carried out to spectral domain and spatial information (si) simultaneously with DBM, obtains the category feature of pixel, more differentiates
Ability is conducive to later period classification.
In another embodiment, the spectrum of the synthesis-spatial feature vector h merging is exactly spectral signature vector
hspeWith spatial feature vector hspaHead and the tail connect, and are stitched together.
In another embodiment, the spectrum of the synthesis-spatial feature vector h length is equal to spectral signature vector hspe
With spatial feature vector hspaThe sum of length.
For the embodiment, it is merely given as comprehensive one of spectrum-spatial feature vector merging mode, this method is same
Sample is suitable for the spectrum-spatial feature vector for the synthesis that other methods merge out.
Following embodiment verifies feature and performance of the DBM model in classification hyperspectral imagery.
Two high-spectral data collection for being widely used as comparison benchmark: Indian Pines and Pavia are applied in experiment
University.Indian Pines data set is the remote sensing of the Indiana northwestern obtained by AVIRIS sensor
Image.The data set includes 145x145 pixel and 200 effective spectral band information.To data concentrate pixel class into
It has gone artificial mark, has contained 16 class atural object classifications.Pavia University data set is obtained by ROSIS-03 sensor
It takes, spatial resolution 1.3m, 103 effective band class informations is remained after eliminating invalid spectral coverage, are manually labelled with 9 types
Other atural object.The sample number of two datasets is as shown in table 1.In an experiment, 10% is randomly selected as training to every a kind of sample
Sample, remaining 90% is used as test sample.
Table 1
In another embodiment, by hidden layer number in experimental verification DBM and the number of nodes of each hidden layer to classification results
Influence.
DBM model is the depth structure with more hidden layers.Hidden layer number is more, then model can extract data more
The expression of higher level of abstraction, result are more advantageous to classifier and carry out discriminant classification.But in practical application, as hidden layer number increases
More, the number of parameters of model can sharply increase.For limited training data, it is easy to over-fitting is caused, to reduce test
The discrimination of data.The interstitial content of each layer also has similar problem: number of nodes excessively easily causes over-fitting, and does not have then less excessively
Sufficiently feature between extraction data.
(1) influence of the hidden layer number to classification results
Hidden layer number is tested respectively are as follows: classification results when 2,3,4,5,6.Each the number of hidden nodes is set as 150 in experiment,
Pre-training and the number of iterations in fine tuning stage are set as 1000 times.Experimental data only with pixel spectroscopic data.Meanwhile it taking
Average accuracy after program operation 10 times is as final result, as shown in Figure 3.
From figure 3, it can be seen that the optimum number of strata of Indian Pines data set is 2 layers, and Pavia University
Optimum number of strata be 3 layers.Therefore illustrate that the hidden layer number of DBM model is associated with the data set of particular problem.
(2) influence of the number of hidden nodes to classification results
In order to determine the Hidden nodes of model, experiment sets 80 for every layer of Hidden nodes, then with 5 for increment gradually
Increase, until 200, counts corresponding classification accuracy rate respectively.In experiment, the number of iterations of pre-training and fine tuning is respectively provided with
It is 1000 times, the hidden layer number of two data is respectively set to its optimal value.Experimental data only with pixel spectroscopic data.As a result
As shown in Figure 4, it can be seen that the optimal Hidden nodes of Indian Pines and Pavia University are respectively 150 and 120.
In general, in the number of hidden nodes and input data number of features there is positive incidence.Work as sample data
Dimension it is bigger when, optimal the number of hidden nodes is also more.
In another embodiment, it demonstrates for DBN, the mistake of DBM divides rate can be rapid with iterative process
It reduces.Fig. 5 (a) show the DBN model to be stacked up by two RBM, from the point of view of DBN angle, it include visible layer V and
Two hidden layer h1、h2。W1, W2It is visible layer and hidden layer h respectively1And two hidden layer h1With h2Between weight connection matrix.DBM
Its structure of model and node connection type such as Fig. 5 (b) are shown, have carried out two o'clock improvement to the learning algorithm of DBN model: (1) having existed
When calculating the activation value of hidden node, while considering the input information of upper and lower two adjacent node layers;(2) it is using
After layer-by-layer greedy pre-training algorithm during DBN pre-training, the total instruction for being based on Mean Field (MF) method is increased
Practice process.
Therefore, in DBM model, the condition probability formula of each node layer is calculated are as follows:
V=(v in formula1, v2..., vn) indicate DBM visible layer knot vector, vi(i=1,2 ..., n) is the value of visible node, and n is
It can be seen that interstitial content, value is equal to the dimension of sample data.With
Indicate that the first and second hidden node of DBM vector, s and m respectively indicate first and second layers of hidden node number, s and m are positive whole
Number.WithRespectively indicate the first hidden node and the second hidden layer section
The activation value of point.ai、bj、ckRespectively indicating the bias of the node of visible layer, the first hidden layer and the second hidden layer, (i, j, k value are same
On).Indicate the connection weight between j-th of node of i-th of node of visible node layer and the first hidden layer.Indicate the first hidden layer
Connection weight between k-th of node of j-th of node and the second hidden layer.FunctionFor sigmoid function.
From (2), formula can be seen that hidden layer h1Activate the hidden layer h of probability by following visible layer and above2Input common group
At.It is this to improve so that DBM model instruct its study using priori knowledge in the training process relative to DBN model
Depth dependency in data.
Due in DBM model-learning algorithm calculate hidden node activation value when, while consider up and down node layer
Input information can be using the classification of sample data as priori knowledge therefore in its pre-training stage.Based on this, in DBM
On the top layer of model, then logically increase a mark layer l, the number of the node layer is equal to sample class number, then (3) formula
Calculate hidden layer h2Activation value modification are as follows:
It is wherein l=(L1, l2..., lq) mark layer vector, ll(l=1,2 ..., q) is the value for marking node layer, and q is sample
This classification number is positive integer.For the connection weight between the 1st node of k-th of node and mark layer of the second hidden layer.
This learning algorithm this training of DBM is become have the training of supervision.
Its shadow to classification hyperspectral imagery is verified by the classification experiments on Pavia University data set
It rings.Model hidden layer number is set as 3 layers in an experiment, the number of hidden nodes 120, the iteration of pre-training process and the total training process of MF
Number is 100.At the end of each iteration of training process, statistics is calculated error and is divided rate by the sample number of mistake classification.Figure
6 show that DBN model (black), DBM model remove MF always training (blue) and DBM model and always train (red) three plus MF
The curve graph that the mistake of kind method divides rate to change with iteration.
It can be seen from the figure that the mistake of DBM model divides rate can be rapid with iterative process for DBN model
When reducing, and always training in DBM training process plus MF, mistake divides rate decline faster.
In another embodiment, by this method and the method based on RBF-SVM model, AE model and DBN model into
Row comparison, to verify the validity of this method.For fair comparison in these experiments, carried out to original high-spectral data
It is unified to retain preceding 4 principal components when PCA processing.When constructing spatial information (si), the unified pixel Size of Neighborhood that is arranged is 5 × 5.?
In deep learning model, the unified hidden layer number that is arranged is 2, and other parameters take optimum value.The number of iterations is disposed as 1000 times.Often
A experiment repeats 10 times, calculates overall accuracy (OA), three Measure Indexes (hundred of mean accuracy (AA) and Kappa coefficient
Divide ratio), classification of assessment result is come with average value.Table 2 shows the mean value and mean square deviation in these three indexs on both data sets.
Table 2
From the results, it was seen that this method all achieves best result in the classificatory overall accuracy of two datasets.
Its reason essentially consists in two o'clock: (1) after the spectral information and spatial information (si) for having merged pixel, DBM can learn between data
Higher level of abstraction dependence so that extracted feature has more discriminating power;(2) each hidden in the training algorithm of DBM
The activation probability of layer receives the input information from two neighboring layer simultaneously, and can apply the category prior information of data.
This feature that DBM is learnt from high-spectral data has more discriminating power.
In the training algorithm of DBM, to increase calculation amount as cost, learning ability is improved.In above-mentioned experiment, DBM's
Average workout times are 1.86 times of DBN, and overall accuracy averagely improves 0.61% on both data sets.But it is testing
On time, DBM ratio DBN only increases by 8.06%.Therefore, comprehensively consider calculation amount and nicety of grading, in classification hyperspectral imagery,
DBM still has use and researching value.
Although embodiment of the present invention is described in conjunction with attached drawing above, the invention is not limited to above-mentioned
Specific embodiments and applications field, above-mentioned specific embodiment are only schematical, directiveness, rather than restricted
's.Those skilled in the art are under the enlightenment of this specification and in the range for not departing from the claims in the present invention and being protected
In the case where, a variety of forms can also be made, these belong to the column of protection of the invention.
Claims (7)
1. a kind of fusion spectrum-spatial information (si) depth Boltzmann machine DBM hyperspectral image classification method, includes the following steps:
S100: to spatial information (si) of its space neighborhood information of each image element extraction as the pixel in original high spectrum image, and
Combine the spectrum-spatial information (si) to form the pixel with its spectrum domain information;
S200: feature extraction is carried out by spectrum-spatial information (si) of the depth Boltzmann machine DBM to the pixel;
S300: classified using logistic regression LR to extracted feature, obtain the category label of each pixel.
2. the method according to claim 1, wherein preferred, step S100 further comprises:
S101: the spatial feature vector h of pixel is constructedspa;
S102: the spectral signature vector h of pixel is extracted in original high spectrum imagespe;
S103: the spectral signature vector h of the pixelspeWith spatial feature vector hspaMerge into comprehensive spectrum-spatial feature to
Measure h.
3. method according to claim 2, wherein step 101 further comprises:
S1001: PCA whitening processing is carried out along spectrum direction to the original high spectrum image, k principal component before retaining;
S1002: intercepting the neighborhood of c × c size of each pixel, and be drawn into one-dimensional vector, and obtaining length is c × c × k sky
Characteristic of field vector hspa;
Wherein, for the value range of k between 1 and l, l is the length of the spectral vector of pixel, and the value range of c arrives each picture for 3
The length and wide maximum value of member.
4. method according to claim 2, wherein step 102 further comprises:
Comprehensive spectrum-spatial feature vector h is input in DBM, feature extraction is carried out.
5. method according to claim 4, the process of the feature extraction refer to using comprehensive spectrum-spatial feature vector h as
The visible layer vector of DBM, inside DBM, obtained hidden layer vector is exactly extracted feature vector after upward reasoning.
6. method according to claim 2, comprehensive spectrum-spatial feature vector h merging is spectrum described in step S103
Feature vector hspeWith spatial feature vector hspaHead and the tail connect, and are stitched together.
7. method according to claim 2, wherein the spectrum of the synthesis-spatial feature vector h length be equal to spectral signature to
Measure hspeWith spatial feature vector hspaThe sum of length.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811616702.7A CN109711470A (en) | 2018-12-27 | 2018-12-27 | Merge spectrum-spatial information (si) DBM hyperspectral image classification method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811616702.7A CN109711470A (en) | 2018-12-27 | 2018-12-27 | Merge spectrum-spatial information (si) DBM hyperspectral image classification method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109711470A true CN109711470A (en) | 2019-05-03 |
Family
ID=66257968
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811616702.7A Pending CN109711470A (en) | 2018-12-27 | 2018-12-27 | Merge spectrum-spatial information (si) DBM hyperspectral image classification method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109711470A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114756823A (en) * | 2022-04-18 | 2022-07-15 | 四川启睿克科技有限公司 | Method for improving pepper spectrum model prediction capability |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106446936A (en) * | 2016-09-06 | 2017-02-22 | 哈尔滨工业大学 | Hyperspectral data classification method for spectral-spatial combined data and oscillogram conversion based on convolution neural network |
US20170173262A1 (en) * | 2017-03-01 | 2017-06-22 | François Paul VELTZ | Medical systems, devices and methods |
US9805255B2 (en) * | 2016-01-29 | 2017-10-31 | Conduent Business Services, Llc | Temporal fusion of multimodal data from multiple data acquisition systems to automatically recognize and classify an action |
CN108898156A (en) * | 2018-05-28 | 2018-11-27 | 江苏大学 | A kind of green green pepper recognition methods based on high spectrum image |
-
2018
- 2018-12-27 CN CN201811616702.7A patent/CN109711470A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9805255B2 (en) * | 2016-01-29 | 2017-10-31 | Conduent Business Services, Llc | Temporal fusion of multimodal data from multiple data acquisition systems to automatically recognize and classify an action |
CN106446936A (en) * | 2016-09-06 | 2017-02-22 | 哈尔滨工业大学 | Hyperspectral data classification method for spectral-spatial combined data and oscillogram conversion based on convolution neural network |
US20170173262A1 (en) * | 2017-03-01 | 2017-06-22 | François Paul VELTZ | Medical systems, devices and methods |
CN108898156A (en) * | 2018-05-28 | 2018-11-27 | 江苏大学 | A kind of green green pepper recognition methods based on high spectrum image |
Non-Patent Citations (2)
Title |
---|
SALAKHUTDINOV R 等: "Deep Boltzmann Machines", 《JOURNAL OF MACHINE LEARNING RESEARCH》 * |
林洲汉: "基于自动编码机的高光谱图像特征提取及分类方法研究", 《中国优秀硕士学位论文全文数据库》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114756823A (en) * | 2022-04-18 | 2022-07-15 | 四川启睿克科技有限公司 | Method for improving pepper spectrum model prediction capability |
CN114756823B (en) * | 2022-04-18 | 2024-06-11 | 四川启睿克科技有限公司 | Method for improving prediction capability of pepper spectrum model |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110443143B (en) | Multi-branch convolutional neural network fused remote sensing image scene classification method | |
CN110399909B (en) | Hyperspectral image classification method based on label constraint elastic network graph model | |
CN106023065B (en) | A kind of tensor type high spectrum image spectral-spatial dimension reduction method based on depth convolutional neural networks | |
CN110348399B (en) | Hyperspectral intelligent classification method based on prototype learning mechanism and multidimensional residual error network | |
CN104751191B (en) | A kind of Hyperspectral Image Classification method of sparse adaptive semi-supervised multiple manifold study | |
CN105760821B (en) | The face identification method of the grouped accumulation rarefaction representation based on nuclear space | |
CN103971123B (en) | Hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL) | |
CN109766858A (en) | Three-dimensional convolution neural network hyperspectral image classification method combined with bilateral filtering | |
CN107451565B (en) | Semi-supervised small sample deep learning image mode classification and identification method | |
CN107239759B (en) | High-spatial-resolution remote sensing image transfer learning method based on depth features | |
CN104616032A (en) | Multi-camera system target matching method based on deep-convolution neural network | |
CN103699874B (en) | Crowd abnormal behavior identification method based on SURF (Speed-Up Robust Feature) stream and LLE (Locally Linear Embedding) sparse representation | |
CN110533077A (en) | Form adaptive convolution deep neural network method for classification hyperspectral imagery | |
CN106529484A (en) | Combined spectrum and laser radar data classification method based on class-fixed multinucleated learning | |
CN104732244A (en) | Wavelet transform, multi-strategy PSO (particle swarm optimization) and SVM (support vector machine) integrated based remote sensing image classification method | |
CN112580480B (en) | Hyperspectral remote sensing image classification method and device | |
CN107194423A (en) | The hyperspectral image classification method of the integrated learning machine that transfinites of feature based random sampling | |
CN110147725A (en) | A kind of high spectrum image feature extracting method for protecting projection based on orthogonal index office | |
CN106529458A (en) | Deep neural network space spectrum classification method for high-spectral image | |
CN109472733A (en) | Image latent writing analysis method based on convolutional neural networks | |
CN112949738A (en) | Multi-class unbalanced hyperspectral image classification method based on EECNN algorithm | |
Yuan et al. | ROBUST PCANet for hyperspectral image change detection | |
CN115496934A (en) | Hyperspectral image classification method based on twin neural network | |
CN106096622A (en) | Semi-supervised Classification of hyperspectral remote sensing image mask method | |
CN107169407A (en) | Hyperspectral image classification method based on joint bilateral filtering and extreme learning machine |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190503 |