CN108491849A - Hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks - Google Patents
Hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks Download PDFInfo
- Publication number
- CN108491849A CN108491849A CN201810243423.4A CN201810243423A CN108491849A CN 108491849 A CN108491849 A CN 108491849A CN 201810243423 A CN201810243423 A CN 201810243423A CN 108491849 A CN108491849 A CN 108491849A
- Authority
- CN
- China
- Prior art keywords
- dimensional
- layer
- spectrum
- dense
- block
- 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
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- 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
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/58—Extraction of image or video features relating to hyperspectral data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/194—Terrestrial scenes using hyperspectral data, i.e. more or other wavelengths than RGB
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (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)
- Multimedia (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of hyperspectral image classification methods based on three-dimensional dense connection convolutional neural networks, have the feature that, include the following steps：Step 1 inputs the three-dimensional cubic data block of high spectrum image；Step 2, characteristic pattern between being composed with three-dimensional dense spectrum block processing three-dimensional cubic data block；Step 3, characteristic pattern obtains compressive features figure between handling spectrum with 3-D transition layer；Step 4 obtains space characteristics figure with three-dimensional dense spatial spectrum block processing compressive features figure；Step 5, space characteristics figure obtain prediction label vector by pond layer, compression layer, dropout layers, full articulamentum；Step 6 determines object function；Prediction label vector substitution object function is obtained the loss of repetitive exercise by step 7；Step 8 optimizes parameter to be optimized according to loss；Step 1 is repeated several times to five, seven, eight in step 9, repeatedly optimizes prediction label vector when parameter to be optimized obtains loss reduction, the i.e. classification results of high spectrum image.
Description
Technical field
The present invention relates to a kind of methods of classification hyperspectral imagery, and in particular to one kind is based on three-dimensional dense connection convolution god
Hyperspectral image classification method through network.
Background technology
The normal image of traditional only a small amount of information only has very narrow visible light wave range.Bloom spectrum sensor generally has
A wave bands up to a hundred, the independent signal absorbed in the wavelength band of each wave band, according to different substances to each wave band
The different feedback signals of spectrum generate corresponding two dimensional image, and the data of all wave bands finally form a multichannel together
Three-dimensional data.Therefore, high spectrum image includes a large amount of information, has many typical cases, is such as used for civil and military neck
High spectrum image target detection in domain.Wherein, classification hyperspectral imagery has very important effect in remote sensing fields.
For classification hyperspectral imagery, machine learning related algorithm and different feature extracting methods are answered by many scholars
For classification hyperspectral imagery.2015, Li propose it is a kind of using local binary pattern (LBPs) extract characteristics of image add
Frame of the upper efficient extreme learning machine (ELM) simple in structure as grader.It should be the experimental results showed that LBP be carried in space characteristics
Take largely effective, and compared with the method based on Support vector machine (SVM), ELM graders are more efficient.However,
Compared to LBP feature extracting methods, complicated spectrum and the spatial information of HSI need more complicated cleverly feature selection approach (Li
W,Chen C,Su H,et al.Local Binary Patterns and Extreme Learning Machine for
Hyperspectral Imagery Classification[J].IEEE Transactions on Geoscience and
Remote Sensing,2015,53(7):3681-3693.).Deng et al. proposes a kind of microtexture (micro based on HSI
Texture HSI taxonomy models).Local acknowledgement's pattern (LRP) is generalized to and indicates HSI in texture enhancing (TE) by the frame,
It is reduced using differentiation locality preserving projections (Discriminated Locality Preserving Projection, DLPP)
The data dimension of HSI.But the frame utilize HSI spectral information, performance can further improve (Deng S, Xu Y,
He Y,et al.A hyperspectral image classification framework and its application
[J].Information Sciences, 2015,299(Supplement C):379-393.).Jimenez et al. will be a kind of
The multi-resolution segmentation algorithm of region-growth is applied in the post-processing of HSI classification, to improve the performance of sorting technique.The party
Method will classify to be combined with segmentation, has been obviously improved classification results, but its major defect is the cost of multi-resolution algorithm optimization
(Jimenez L I,PlazaA,Ayma V A,et al. Segmentation as postprocessing for
hyperspectral image classification[C]. International Conference on Computer
as a Tool,IEEE EUROCON 2015, September 8,2015-September 11,2015,2015.)。2016
Year, Huang et al. merges the nicety of grading that can effectively improve HSI with spatial information in conjunction with spectrum, it is proposed that a kind of based on KNN's
Hyperspectral image classification method.This method improves the EO-1 hyperion pixel classifications probability obtained by SVM using KNN filtering algorithms
Figure carries out the Federated filter of image, to which the spectrum and sky of HSI be utilized simultaneously in conjunction with the value and space coordinate of different pixels
Between information (Huang K, Li S, Kang X, et al.SpectralSpatial Hyperspectral Image
Classification Based on KNN[J].Sensing and Imaging,2016,17(1): 1-13.)。
In recent years, it compares and above traditional machine learning related algorithm such as SVM, ELM etc., depth convolutional neural networks
(DCNN) in computer vision have more dominant position, conventional machines learning algorithm in the visual identity of big data (such as
ILSVRC it) is difficult to compared with depth convolutional network, so that be that shallow-layer learns (Shallow Learning) by some persons.2016
Year, Chen et al. proposes a kind of regularization depth characteristic (feature extraction, FE) based on convolutional neural networks
Extracting method, and establish spectral space feature of FE models of the 3D based on CNN for extracting HSI.In order to further carry
High-performance proposes a kind of virtual sample Enhancement Method.The paper illustrates the potential of deep learning, for further research, establishes
Direction (] Chen Y, Jiang H, Li C, et al.Deep Feature Extraction and Classification
of Hyperspectral Images Based on Convolutional Neural Networks[J].IEEE
Transactions on Geoscience and Remote Sensing, 2016,54(10):6232-6251.)。2017
Year, it is different from above-mentioned et al. the spectrum for needing actively extract in HSI and spatial information, Li et al. people without actively extracting the feature of HSI,
But directly using original three-dimensional high level data as input, it is proposed that a kind of 3D convolutional neural networks frames for HSI classification
Frame.The frame does not depend on any pretreatment or post-processing, can but extract effective spectral space assemblage characteristic.In addition, the frame
Frame is compared with other deep learnings, and required parameter is less, and model is lighter, it is easier to training (Li Y, Zhang H, Shen
Q.Spectral-spatial classification of hyperspectral imagery with 3D
convolutional neural network[J].Remote Sensing,2017,9(1).).Shi based on extraction HSI three
Dimensional feature is largely effective property for HSI classification problems, it is proposed that a kind of three-dimensional multiresolution wavelet convolutional network (3D-
MWCNNs) HSI to be used to classify.3D-MWCNNs is combined on the basis of CNNs frames with multiresolution analysis theory, adaptively
Three-dimensional feature is extracted from different scale and different depth in ground.The spectrum and space characteristics of the model extraction more have judgement index, to
HSI niceties of grading can be improved.But the training time needed for the model is far more than other sorting techniques (Shi C, Pun C-M.3D
multi-resolution wavelet convolutional neural networks forhyperspectral image
classification[J].Information Sciences,2017,420: 49-65.).Zhong et al. proposes one kind
Spectral space remnants networks (spectral-spatial residual network, SSRN) end to end.The network utilizes
Spectrum and spatial redundancy block (residual blocks) learn to have to sentence from the abundant spectral signature and space background in HSI
The feature of other power, and residual block connects other convolutional layers by identical mapping (identity mapping).And reinforce every
BN normalization (Batch normalization) on layer convolutional layer.Therefore, compared with other methods, there is high classification essence
Degree.But compared with other methods, model training overlong time (Zhong Z, Li J, Luo Z, et al.Spectral-
Spatial Residual Network for Hyperspectral Image Classification:A 3-D Deep
Learning Framework[J],2017.)。
To sum up, have in hyperspectral image classification method, there is complicated feature extraction engineering in conventional machines learning method,
And state-of-the-art level is not achieved in nicety of grading.And depth convolutional network and its extended network have very long model training mostly
Time.
Invention content
The present invention is to carry out to solve the above-mentioned problems, and it is an object of the present invention to provide a kind of based on three-dimensional dense connection convolution
The hyperspectral image classification method of neural network
The present invention provides a kind of hyperspectral image classification methods based on three-dimensional dense connection convolutional neural networks, have
Such feature, includes the following steps：Step 1 inputs the three-dimensional cubic data block of the multichannel of high spectrum image；Step 2,
Three-dimensional cubic data block is handled by three-dimensional dense spectrum block, the spectrum between the multichannel of study extraction high spectrum image
Between feature composed between characteristic pattern；Step 3, by 3-D transition layer, characteristic pattern is handled to obtain compressive features figure between spectrum；
Step 4 is handled compressive features figure by three-dimensional dense spatial spectrum block, the space characteristics of study extraction high spectrum image
Obtain space characteristics figure；Step 5, by space characteristics figure successively by pond layer, compression layer, dropout layers and full articulamentum
Obtain a prediction label vector；Step 6 determines object function to be optimized；Prediction label vector is substituted into mesh by step 7
In scalar functions, the loss of repetitive exercise is obtained；Step 8, according to the loss of repetitive exercise, to three-dimensional dense connection convolutional Neural
The parameter to be optimized of network model optimizes；Step 1 is repeated several times to five and Step 7: eight, to be optimized in step 9
Parameter carries out prediction label vector when repeatedly optimization obtains the loss reduction of repetitive exercise, as the classification knot of high spectrum image
Fruit.
Provided by the invention based in the hyperspectral image classification method of three-dimensional dense connection convolutional neural networks, may be used also
To have the feature that：Wherein, the choosing method of three-dimensional cubic data block is：To with the height that L channel sized is H × W
Spectrum picture is denoted as x from the three-dimensional cubic data block for choosing r × r × L sizes in its original pixels centered on a pixel(r ×r×L), the spatial image size of r × r expression three-dimensional cubic data blocks.
Provided by the invention based in the hyperspectral image classification method of three-dimensional dense connection convolutional neural networks, may be used also
To have the feature that：Wherein, prediction label vector y~dimension be C,Dense
() indicates three-dimensional dense connection convolutional neural networks model；C is high spectrum image classification number to be sorted；δ is ginseng to be optimized
Number；The cross entropy loss function L of high spectrum imagesFor,M indicates that specification layer is included
Sample size；Xi indicates i-th layer of input feature vector of full articulamentum, it belongs to yiClass；WjIt indicates in last full articulamentum
Weight W jth row；B indicates bias term；The center loss function L of high spectrum imagecFor,CyiTable
Show yiThe depth characteristic center of class, object function F are cross entropy loss function LsWith center loss function LcThe sum of minimum
Value, i.e.,
Provided by the invention based in the hyperspectral image classification method of three-dimensional dense connection convolutional neural networks, may be used also
To have the feature that：Wherein, further include before step 2：Goal pels cubic number is chosen from three-dimensional cubic data block
According to obtaining the n characteristic pattern that size is r × r × b, input of these characteristic patterns as the dense spectrum block of three-dimensional is denoted as
In subscript 1 indicate that the data are located in three-dimensional dense spectrum block, it is in the block that subscript 0 indicates that the data are located at three-dimensional dense spectrum
Initial position.
Provided by the invention based in the hyperspectral image classification method of three-dimensional dense connection convolutional neural networks, may be used also
To have the feature that：Wherein, three-dimensional dense spectrum block includes being connected with dense connection type with three-dimensional dense space block
The compound convolutional layer of multi-layer three-dimension, every layer of compound convolutional layer includes sequentially connected specification layer, active coating and Three dimensional convolution layer,
The size of the convolution kernel of the Three dimensional convolution layer of three-dimensional dense spectrum block is 1 × 1 × d, and the quantity of convolution kernel is k, three-dimensional dense sky
Between the size of convolution kernel of Three dimensional convolution layer of block be a × a × g, the quantity of convolution kernel is k, and a × a is the size of spatial image,
D, g is the quantity in channel.
Provided by the invention based in the hyperspectral image classification method of three-dimensional dense connection convolutional neural networks, may be used also
To have the feature that：Wherein, the compound convolutional layer of multi-layer three-dimension of three-dimensional dense spectrum block is denoted as D1(), three-dimensional dense light
The output of the 1st layer of compound convolutional layer of three-dimensional is in spectrum blockL layers of three-dimensional are compound in three-dimensional dense spectrum block
The output of convolutional layer isThe l layer compound convolutional layer of three-dimensional of three-dimensional dense spectrum block it is defeated
The characteristic pattern gone out is characteristic pattern between composing, the quantity of characteristic pattern between spectrumFor
Provided by the invention based in the hyperspectral image classification method of three-dimensional dense connection convolutional neural networks, may be used also
To have the feature that：Wherein, further include before step 3 between spectrum the channel of characteristic pattern merge to obtain kmBetween a spectrum
Characteristic pattern.
Provided by the invention based in the hyperspectral image classification method of three-dimensional dense connection convolutional neural networks, may be used also
To have the feature that：Wherein, 3-D transition layer includes sequentially connected specification layer and three layers of Three dimensional convolution layer, three-dimensional mistake
The size for crossing the convolution kernel of first Three dimensional convolution layer of layer is 1 × 1 × 1, and the quantity of convolution kernel is θ × kmIt is a, this first
Three dimensional convolution layer is for changing the number of channels of characteristic pattern between spectrum, the structure of second and third Three dimensional convolution layer of 3-D transition layer
It is identical, it is used to remold the shape of characteristic pattern between spectrum, reduces the space size of the data block of characteristic pattern between composing.
Provided by the invention based in the hyperspectral image classification method of three-dimensional dense connection convolutional neural networks, may be used also
To have the feature that：Wherein, the compound convolutional layer of multi-layer three-dimension of three-dimensional dense space block is denoted as D2(), three-dimensional dense sky
Between in block the output of l layers of three-dimensional compound convolutional layer beThe l of three-dimensional dense space block
The characteristic pattern of the output of the three-dimensional compound convolutional layer of layer is space characteristics figure.
The effect of invention
According to the hyperspectral image classification method according to the present invention based on three-dimensional dense connection convolutional neural networks, because
For using the three-dimensional cubic data block of high spectrum image as input, EO-1 hyperion directly is automatically extracted using three-dimensional dense spectrum block
Characteristic pattern between feature is composed between abundant spectrum in image, by 3-D transition layer, characteristic pattern is handled and is compressed between spectrum
Characteristic pattern obtains space characteristics figure using three-dimensional dense space block to automatically extract space characteristics abundant in high spectrum image,
So the network of the dense connection convolutional neural networks of three-dimensional involved by this method is deeper without there is the problem of gradient disappearance,
Significantly more efficient feature can be utilized, and strengthen the transmission of feature, and then improve the precision of classification, while shorten instruction
Practice the time.
In addition, repeatedly being optimized to treat Optimal Parameters by the loss of repetitive exercise for this method, that is, use dynamic
The learning rate of state, also uses dropout, these make the convergence rate of three-dimensional dense connection convolutional neural networks accelerate, only
The iteration of fewer number, which need to be passed through, can reach best nicety of grading.
Description of the drawings
Fig. 1 is the hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks in the embodiment of the present invention
Flow chart；
Fig. 2 is the structural representation of the dense spectrum block of three-dimensional with three layers of compound convolutional layer of three-dimensional in the embodiment of the present invention
Figure；
Fig. 3 is the structural schematic diagram of 3-D transition layer in the embodiment of the present invention；
Fig. 4 is the structural representation of the dense space block of three-dimensional with three layers of compound convolutional layer of three-dimensional in the embodiment of the present invention
Figure；
Fig. 5 is the relational graph of object function and iterations in the embodiment of the present invention；
Fig. 6 is in the embodiment of the present invention using the classification hyperspectral imagery based on three-dimensional dense connection convolutional neural networks
The visualization result that method (3D-DenseNet) classifies to PaviaUniversity (UP) data；And
Fig. 7 is in the embodiment of the present invention using the classification hyperspectral imagery based on three-dimensional dense connection convolutional neural networks
The visualization result that method (3D-DenseNet) classifies to Indiana Pines (IN) data.
Specific implementation mode
It is real below in order to make the technical means, the creative features, the aims and the efficiencies achieved by the present invention be easy to understand
Example combination attached drawing is applied to the present invention is based on the hyperspectral image classification method works of three-dimensional dense connection convolutional neural networks specifically to explain
It states.
Fig. 1 is the hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks in the embodiment of the present invention
Flow chart.
As shown in Figure 1, the hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks includes following step
Suddenly：
Step S1：
Input the data set of the original pixels of high spectrum image and corresponding true tag.
{(x(r×r×L), gth), y }
x(r×r×L)Indicate three-dimensional cubic data block, it is original from its to the high spectrum image with L channel sized for H × W
The three-dimensional cubic data block for choosing r × r × L sizes in pixel centered on a pixel, is denoted as x(r×r×L), wherein r × r tables
Show the spatial image size of three-dimensional cubic data block.The true tag (ground truth) of high spectrum image is denoted as input gth,
gth∈R2.For inputting to (x(r×r×L), gth), corresponding output is y, y ∈ R2。
Step S2：
Preparation standards are carried out to three-dimensional cubic data block and obtain standardized data.It is obtaining as a result, for every row/
All data are all gathered near 0 for each column, variance 1.
Step S3：
Standardized data is filled to expand with zero in three dimensions and obtains filling data.
Step S4：
Data distribution is carried out to filling data, process is as follows：One is pressed with pixel quantity of the high spectrum image in spatial domain
Fixed pro rate is training dataset, validation data set and test data set.The single channel image of EO-1 hyperion shares TN picture
Vegetarian refreshments, that is, TN index, wherein TN=r × r, r are the size of spatial domain.By the rope of markd pixel in true tag
Draw and extract, makes pixel index list, be denoted as LN.Model training after for convenience, verification and test, by pixel rope
Draw list LN and is randomly divided into trained indexed set LN according to a certain percentagetrain, verification indexed set LNvalWith test indices collection LNtest。
Step S5：
Goal pels cubic data is chosen, process is as follows：By goal pels centered on the position in spatial domain from filling
A certain size small cubic block is chosen in data.According to training indexed set LNtrain, verification indexed set LNvalWith test indices collection
LNtest, filling data block is assigned as training data block collection Xtrain, verify data block collection XvalWith test data block collection Xtest。
Step S6：
Fig. 2 is the structural representation of the dense spectrum block of three-dimensional with three layers of compound convolutional layer of three-dimensional in the embodiment of the present invention
Figure.
As shown in Fig. 2, three-dimensional dense spectrum block includes the compound convolutional layer of multi-layer three-dimension connected with dense connection type.
Three-dimensional dense spectrum block has three layers of compound convolutional layer of three-dimensional in the present embodiment.The compound convolutional layer of every layer of three-dimensional includes being sequentially connected
Specification layer, active coating and Three dimensional convolution layer.Wherein, the activation primitive of active coating is parametrization amending unit；Three dimensional convolution
The size of the convolution kernel of layer is 1 × 1 × d, and d is the quantity in channel, and the quantity of convolution kernel is k.One layer of three-dimensional compound convolutional layer
K characteristic pattern can be generated, the growth rate of neural network is referred to as.
By three-dimensional dense spectrum block to training data block collection XtrainIt is handled, study extraction high spectrum image leads to more
Feature between spectrum between road, process are as follows：
Training data block collection XtrainIt first passes around Three dimensional convolution and generates the n characteristic pattern that size is r × r × b, these features
Scheme, as the input of three-dimensional dense spectrum block, to be denoted asIn subscript 1 indicate that the data are located at three-dimensional dense spectrum block
In, subscript 0 indicates that the data are located at three-dimensional dense spectrum initial position in the block；Three-dimensional compound convolutional layer is denoted as D1(), the 1st
The output of the three-dimensional compound convolutional layer of layer is The output of l layers of three-dimensional compound convolutional layer is
The characteristic pattern of the output of l layers of three-dimensional compound convolutional layer of three-dimensional dense spectrum block is study extraction spectrum picture
Characteristic pattern between the spectrum obtained after feature between spectrum.In three-dimensional dense spectrum block, the input feature vector figure of the compound convolutional layer of every layer of three-dimensional
Size be r × r × b, export characteristic pattern size be also r × r × b, export characteristic pattern quantity be k, but export
The quantity of characteristic pattern is deepened and linear increase with the number of plies of three-dimensional compound convolutional layer, therefore, the quantity of characteristic pattern between spectrum
For
Step S7：
It merges the channel of characteristic pattern between spectrum to obtain kmCharacteristic pattern between the spectrum that a size is r × r × b.
Step S8：
Fig. 3 is the structural schematic diagram of 3-D transition layer in the embodiment of the present invention.
As shown in figure 3,3-D transition layer includes sequentially connected specification layer and three layers of Three dimensional convolution layer.Wherein, first
The size of the convolution kernel of layer Three dimensional convolution layer is 1 × 1 × 1, and the quantity of convolution kernel is θ × kmIt is a, 0 θ≤1 ＜；Second and third layer three
The structure for tieing up convolutional layer is identical, is used to remold the shape of characteristic pattern between spectrum.
By 3-D transition layer to kmCharacteristic pattern is handled between the spectrum that a size is r × r × b, and process is as follows：
kmCharacteristic pattern first passes around specification layer and carries out batch standardization between the spectrum that a size is r × r × b, then passes through first layer
Three dimensional convolution layer reduces the number of channels of characteristic pattern between spectrum, is remolded finally by second and third layer of Three dimensional convolution layer special between spectrum
The shape of figure is levied, the space size for reducing data block obtains compressive features figure.The number of compressive features figure is θ × kmIt is a.
Step S9：
Fig. 3 is the structural representation of the dense space block of three-dimensional with three layers of compound convolutional layer of three-dimensional in the embodiment of the present invention
Figure.
As shown in figure 3, three-dimensional dense space block includes the compound convolutional layer of multi-layer three-dimension connected with dense connection type.
Three-dimensional dense space block has three layers of compound convolutional layer of three-dimensional in the present embodiment.The compound convolutional layer of every layer of three-dimensional includes being sequentially connected
Specification layer, active coating and Three dimensional convolution layer.Wherein, the activation primitive of active coating is parametrization amending unit；Three dimensional convolution
The size of the convolution kernel of layer is a × a × g, and a × a is the size of spatial image, and g is the quantity in channel, and the quantity of convolution kernel is k.
One layer of three-dimensional compound convolutional layer can generate k characteristic pattern.
Compressive features figure is handled by three-dimensional dense space block, between the multichannel of study extraction high spectrum image
Spectrum between feature, process is as follows：
Using compressive features figure as the input of three-dimensional dense space block, it is denoted asIn subscript 2 indicate the data bit
In three-dimensional dense space block, subscript 0 indicates that the data are located at three-dimensional dense space initial position in the block；Three-dimensional dense space
The compound convolutional layer of multi-layer three-dimension of block is denoted as D2(), in three-dimensional dense space block the output of the l layer compound convolutional layer of three-dimensional be
While the characteristic pattern of the output of l layers of three-dimensional compound convolutional layer of three-dimensional dense space block is that preservation network is gently narrow
The space characteristics figure obtained after the space characteristics of study extraction high spectrum image.In three-dimensional dense space block, every layer of three-dimensional is multiple
Close convolutional layer input feature vector figure size be r × r × b, export characteristic pattern size be also r × r × b, output feature
The quantity of figure is k, exports quantity linear increase with the number of plies intensification of three-dimensional compound convolutional layer of characteristic pattern, therefore,
The quantity of space characteristics figureFor
Step S10：
Space characteristics figure obtains a prediction label by pond layer, compression layer, dropout layers and full articulamentum successively
Vector.
Prediction label vectorDimension be C,
Dense () indicates three-dimensional dense connection convolutional neural networks model；C is high spectrum image classification to be sorted
Number；δ is parameter to be optimized.
Step S11：
The classification of high spectrum image belongs to classification problem, for such issues that generally use cross entropy (cross-
Entropy) loss function carrys out the difference between predictive metrics value and actual value, and is come to object function function to be optimized with this
It optimizes.Again because classification hyperspectral imagery is more discriminant classifications, therefore softmax is used to return, obtains high spectrum image
Cross entropy loss function LsFor,
Wherein, m indicates the sample size that specification layer is included；Xi indicates i-th layer of input feature vector of full articulamentum, it belongs to
In yiClass；WjIndicate the jth row of the weight W in last full articulamentum；B indicates bias term.
But with cross entropy loss function LsCarry out optimization object function, will produce a problem, i.e., spacing is larger in class.It is each
As soon as a class all safeguards the center of a class, if the feature of the sample is punished too far with class center, here it is on characteristic layer
The heart loses.The center loss function L of high spectrum imagecFor,
Wherein, CyiIndicate yiThe depth characteristic center of class.
To sum up, object function F to be optimized is cross entropy loss function LsWith center loss function LcThe sum of minimum value,
I.e.
Step S12：
Prediction label vector is substituted into object function F, the loss of repetitive exercise is obtained.
Step S13：
According to the loss of repetitive exercise, the parameter δ to be optimized of three-dimensional dense connection convolutional neural networks model is carried out excellent
Change.
Step S14：
Step S1 to S10 and step S12, S13 is repeated several times, treating Optimal Parameters δ progress, repeatedly optimization obtains iteration
Prediction label vector when trained loss reduction, the as classification results of high spectrum image.
Three-dimensional dense connection convolution god in hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks
Convergence verification process through network model is as follows：
The each three-dimensional dense spectrum compound convolutional layer of three-dimensional in the block of experimental setup is 3 layers, each three-dimensional dense space block
In the compound convolutional layer of three-dimensional be 3 layers, the growth rate k of neural network is set as 12, optimizer RMSprop, time of repetitive exercise
Number is 200.Using dynamical learning rate, initial learning rate is 0.0003, and model is not restrained in 10 repetitive exercises, then learning rate
Halve.Target is from training data block collection XtrainMiddle study carries out classification marker to three-dimensional cubic data.
Fig. 5 is the relational graph of object function and iterations in the embodiment of the present invention.
As shown in figure 5, successively increased as iterations have from 0~200, the functional value of object function by 1.94169 by
It decrescence as low as 0.01976 and gradually levels off to 0, shows the dense connection convolutional neural networks model tool of three-dimensional according to the present invention
There is convergence.
High spectrum image is carried out using the hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks
The result of classification is as follows：
Hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks uses PaviaUniversity
Data and Indiana Pines data.
PaviaUniversity data are obtained by ROSIS sensors are in Italy Pavia sub- in 2003.Original number
It is according to size, spatial resolution reaches 1.3m.9 class atural objects and its details are shared in PaviaUniversity data, such as
Shown in table 1.
Table 1 is the classification information of PaviaUniversity data in the embodiment of the present invention
Indiana Pines data were obtained in 1996 in northern state of Indiana western part by AVIRIS spectrometers.Indian
16 kinds of atural objects and its detailed classification information are shared in P data, as shown in table 2.
Table 2 is the classification information of PIndiana Pines data in the embodiment of the present invention
Hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks is in Pavia University numbers
According to randomly select 10 groups of training datas in IndianaPines data and tested, experimental result is with ' average ± variance '
Form provides.Allocation of computer is 32GB memories, 1080TiGPU.This experiment using overall accuracy (overall accuracy,
), OA mean accuracy (average accuracy, AA) and Kappa coefficients (k) and every a kind of specific precision are as classification results
The index of correlation of precision.Size, PaviaUniversity data are inputted using identical 9 × 9 space (r=9) on data set
Training sample ratio with Indiana Pines data is respectively 10% and 20%.
From the point of view of the precision in table 3, this method (3D-DenseNet) to Pavia University (UP) data and
When Indiana Pines (IN) data are classified, overall accuracy (overall accuracy, OA), mean accuracy
Reach 99% or more precision on (average accuracy, AA) and Kappa coefficients (k).
Table 3 is in the embodiment of the present invention using the classification hyperspectral imagery based on three-dimensional dense connection convolutional neural networks
Method (3D-DenseNet) is respectively to Pavia University (UP)
The result that data and IndianaPines (IN) data are classified
Fig. 6 is in the embodiment of the present invention using the classification hyperspectral imagery based on three-dimensional dense connection convolutional neural networks
The visualization result that method (3D-DenseNet) classifies to Pavia University (UP) data.
Fig. 7 is in the embodiment of the present invention using the classification hyperspectral imagery based on three-dimensional dense connection convolutional neural networks
The visualization result that method (3D-DenseNet) classifies to IndianaPines (IN) data.
As shown in fig. 6, the real image on a certain wave band of Pavia University data is the parts Fig. 6-a, really
Label figure is the parts Fig. 6-b and the result using the hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks
Figure is the parts Fig. 6-c.Real image on a certain wave band of Indiana Pines data is the parts Fig. 7-a, authentic signature figure
For the parts Fig. 7-b and use the result figure of the hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks for figure
The parts 7-c.From Fig. 6,7 it can be seen that the classification results figure and authentic signature figure of this method are almost the same.
The effect of embodiment
The hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks involved by the present embodiment,
Because directly automatically extracting bloom using three-dimensional dense spectrum block using the three-dimensional cubic data block of high spectrum image as input
Characteristic pattern between feature is composed between abundant spectrum in spectrogram picture, by 3-D transition layer, characteristic pattern is handled and is pressed between spectrum
Contracting characteristic pattern obtains space characteristics using three-dimensional dense space block to automatically extract space characteristics abundant in high spectrum image
Figure, thus involved by this method three-dimensional it is dense connection convolutional neural networks network it is deeper without occur gradient disappearance ask
Topic significantly more efficient can utilize feature, and strengthen the transmission of feature, and then improve the precision of classification, shorten simultaneously
Training time.
In addition, repeatedly being optimized to treat Optimal Parameters by the loss of repetitive exercise for this method, that is, use dynamic
The learning rate of state, also uses dropout, these make the convergence rate of three-dimensional dense connection convolutional neural networks accelerate, only
The iteration of fewer number, which need to be passed through, can reach best nicety of grading.
Further, the size of the convolution kernel of the first layer Three dimensional convolution layer of 3-D transition layer is 1 × 1 × 1, and image exists
There are local correlations, the process of convolution is a kind of extraction to local correlations, and the volume that size is 1 × 1 × 1 in spatial domain
Product does not extract any feature, because it gives no thought to pixel and other pixel relationships of periphery.But in fact, size be 1 × 1 ×
1 convolution carries out linear combination or information integration to each pixel on different channels, preserve compose between characteristic pattern r ×
While the size of r × b is constant, it can freely change the quantity in the channel of characteristic pattern between composing.
Further, three-dimensional dense spectrum block includes the multilayer connected with dense connection type with three-dimensional dense space block
Three-dimensional compound convolutional layer, this dense convolution connection type make the networks of three-dimensional dense connection convolutional neural networks more it is deep but not
The gradient disappearance problem that will appear so that network is narrower and strengthens the transmission of the feature between convolutional layer, and then improves classification essence
Degree.
The above embodiment is the preferred case of the present invention, is not intended to limit protection scope of the present invention.
Claims (9)
1. a kind of hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks, using three-dimensional dense connection
Convolutional neural networks model classify to high spectrum image, which is characterized in that include the following steps：
Step 1 inputs the three-dimensional cubic data block of the multichannel of high spectrum image；
Step 2 is handled the three-dimensional cubic data block by three-dimensional dense spectrum block, and the EO-1 hyperion is extracted in study
Characteristic pattern between feature is composed between spectrum between the multichannel of image；
Step 3, by 3-D transition layer, characteristic pattern is handled to obtain compressive features figure between the spectrum；
Step 4 is handled the compressive features figure by three-dimensional dense spatial spectrum block, and the high-spectrum is extracted in study
The space characteristics of picture obtain space characteristics figure；
The space characteristics figure is obtained one by step 5 by pond layer, compression layer, dropout layers and full articulamentum successively
A prediction label vector；
Step 6 determines object function to be optimized；
The prediction label vector is substituted into the object function, obtains the loss of repetitive exercise by step 7；
Step 8, according to the loss of the repetitive exercise, to the parameter to be optimized of three-dimensional dense connection convolutional neural networks model
It optimizes；
Step 9 is repeated several times step 1 to five and Step 7: eight, carries out repeatedly optimization to the parameter to be optimized and obtain institute
State the prediction label vector when loss reduction of repetitive exercise, the classification results of the as described high spectrum image.
2. the hyperspectral image classification method according to claim 1 based on three-dimensional dense connection convolutional neural networks,
It is characterized in that：
Wherein, the choosing method of the three-dimensional cubic data block is：To with L channel sized be H × W high spectrum image,
From the three-dimensional cubic data block for choosing r × r × L sizes in its original pixels centered on a pixel, it is denoted as x(r×r×L),
R × r indicates the spatial image size of the three-dimensional cubic data block.
3. the hyperspectral image classification method according to claim 1 based on three-dimensional dense connection convolutional neural networks,
It is characterized in that：
Wherein, the prediction label vectorDimension be C,
Dense () indicates three-dimensional dense connection convolutional neural networks model；C is high spectrum image classification to be sorted
Number；δ is parameter to be optimized；
The cross entropy loss function L of the high spectrum imagesFor,
M indicates the sample size that specification layer is included；Xi indicates i-th layer of input feature vector of the full articulamentum, it belongs to yi
Class；WjIndicate the jth row of the weight W in the last full articulamentum；B indicates bias term；
The center loss function L of the high spectrum imagecFor,
CyiIndicate yiThe depth characteristic center of class,
The object function F is cross entropy loss function LsWith center loss function LcThe sum of minimum value, i.e.,
4. the hyperspectral image classification method according to claim 1 based on three-dimensional dense connection convolutional neural networks,
It is characterized in that：
Wherein, further include before step 2：Goal pels cubic data is chosen from the three-dimensional cubic data block, obtains size
For the n characteristic pattern of r × r × b, input of these characteristic patterns as the dense spectrum block of the three-dimensional is denoted asIn under
Mark 1 indicates that the data are located in three-dimensional dense spectrum block, and subscript 0 indicates that the data are located at three-dimensional dense spectrum start bit in the block
It sets.
5. the hyperspectral image classification method according to claim 1 based on three-dimensional dense connection convolutional neural networks,
It is characterized in that：
Wherein, the dense spectrum block of the three-dimensional includes the multilayer connected with dense connection type with the dense space block of the three-dimensional
Three-dimensional compound convolutional layer,
Every layer of compound convolutional layer includes sequentially connected specification layer, active coating and Three dimensional convolution layer,
The size of the convolution kernel of the Three dimensional convolution layer of the dense spectrum block of three-dimensional is 1 × 1 × d, and the quantity of convolution kernel is k,
The size of the convolution kernel of the Three dimensional convolution layer of the dense space block of three-dimensional is a × a × g, and the quantity of convolution kernel is k,
A × a is the size of spatial image, and d, g are the quantity in channel.
6. the hyperspectral image classification method according to claim 5 based on three-dimensional dense connection convolutional neural networks,
It is characterized in that：
Wherein, three-dimensional compound convolutional layer described in the multilayer of the dense spectrum block of the three-dimensional is denoted as D1(),
The output of the 1st layer of compound convolutional layer of three-dimensional is in the dense spectrum block of three-dimensional
The output of the l layers of compound convolutional layer of three-dimensional is in the dense spectrum block of three-dimensional
The characteristic pattern of the output of the l layers of compound convolutional layer of three-dimensional of the dense spectrum block of three-dimensional is feature between the spectrum
Scheme, the quantity of characteristic pattern between the spectrumFor
7. the hyperspectral image classification method according to claim 1 based on three-dimensional dense connection convolutional neural networks,
It is characterized in that：
Wherein, step 3 further includes that the channel of the characteristic pattern between the spectrum merges to obtain k beforemCharacteristic pattern between a spectrum.
8. the hyperspectral image classification method according to claim 1 based on three-dimensional dense connection convolutional neural networks,
It is characterized in that：
Wherein, the 3-D transition layer includes sequentially connected specification layer and three layers of Three dimensional convolution layer,
The size of the convolution kernel of the first layer Three dimensional convolution layer of the 3-D transition layer is 1 × 1 × 1, the quantity of convolution kernel be θ ×
kmIt is a, the first layer Three dimensional convolution layer for changing characteristic pattern between the spectrum number of channels,
The structure of second and third layer of Three dimensional convolution layer of the 3-D transition layer is identical, the shape for remolding characteristic pattern between the spectrum
Shape reduces the space size of the data block of characteristic pattern between the spectrum.
9. the hyperspectral image classification method according to claim 1 based on three-dimensional dense connection convolutional neural networks,
It is characterized in that：
Wherein, three-dimensional compound convolutional layer described in the multilayer of the dense space block of the three-dimensional is denoted as D2(),
The output of the l layers of compound convolutional layer of three-dimensional is in the dense space block of three-dimensional
The characteristic pattern of the output of the l layers of compound convolutional layer of three-dimensional of the dense space block of three-dimensional is the space characteristics
Figure.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810243423.4A CN108491849B (en) | 2018-03-23 | 2018-03-23 | Hyperspectral image classification method based on three-dimensional dense connection convolution neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810243423.4A CN108491849B (en) | 2018-03-23 | 2018-03-23 | Hyperspectral image classification method based on three-dimensional dense connection convolution neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108491849A true CN108491849A (en) | 2018-09-04 |
| CN108491849B CN108491849B (en) | 2021-09-28 |
Family
ID=63319511
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810243423.4A Active CN108491849B (en) | 2018-03-23 | 2018-03-23 | Hyperspectral image classification method based on three-dimensional dense connection convolution neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108491849B (en) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109784347A (en) * | 2018-12-17 | 2019-05-21 | 西北工业大学 | Image classification method based on multiple dimensioned dense convolutional neural networks and spectrum attention mechanism |
| CN109903255A (en) * | 2019-03-04 | 2019-06-18 | 北京工业大学 | A kind of high spectrum image Super-Resolution method based on 3D convolutional neural networks |
| CN109978041A (en) * | 2019-03-19 | 2019-07-05 | 上海理工大学 | A kind of hyperspectral image classification method based on alternately update convolutional neural networks |
| CN110009015A (en) * | 2019-03-25 | 2019-07-12 | 西北工业大学 | EO-1 hyperion small sample classification method based on lightweight network and semi-supervised clustering |
| CN110263914A (en) * | 2019-05-29 | 2019-09-20 | 朗坤智慧科技股份有限公司 | Deep neural network structure improved method and system based on dense connection |
| CN110852369A (en) * | 2019-11-06 | 2020-02-28 | 西北工业大学 | Hyperspectral image classification method combining 3D/2D convolutional network and adaptive spectrum unmixing |
| CN110852227A (en) * | 2019-11-04 | 2020-02-28 | 中国科学院遥感与数字地球研究所 | Hyperspectral image deep learning classification method, device, equipment and storage medium |
| CN111077093A (en) * | 2020-01-10 | 2020-04-28 | 安徽理工大学 | Method and device for quickly detecting coal gangue based on multispectral technology |
| CN111222453A (en) * | 2020-01-03 | 2020-06-02 | 武汉大学 | Remote sensing image change detection method based on dense connection and geometric structure constraint |
| CN111797941A (en) * | 2020-07-20 | 2020-10-20 | 中国科学院长春光学精密机械与物理研究所 | Image classification method and system carrying spectral information and spatial information |
| CN111950714A (en) * | 2020-08-24 | 2020-11-17 | 重庆市云迈科技有限公司 | Energy spectrum CT image domain material identification method based on 3D full convolution neural network |
| CN112052755A (en) * | 2020-08-24 | 2020-12-08 | 西安电子科技大学 | Semantic convolution hyperspectral image classification method based on multi-path attention mechanism |
| CN112116563A (en) * | 2020-08-28 | 2020-12-22 | 南京理工大学 | Hyperspectral image target detection method and system based on spectral dimension and space cooperation neighborhood attention |
| CN112183399A (en) * | 2020-09-30 | 2021-01-05 | 珠海欧比特宇航科技股份有限公司 | Hyperspectral image classification method and medium based on separable convolution and dense connection |
| CN112633401A (en) * | 2020-12-29 | 2021-04-09 | 中国科学院长春光学精密机械与物理研究所 | Hyperspectral remote sensing image classification method, device, equipment and storage medium |
| CN112950632A (en) * | 2021-04-18 | 2021-06-11 | 吉林大学 | Coal quality detection method based on hyperspectral imaging technology and convolutional neural network |
| CN112966598A (en) * | 2021-03-04 | 2021-06-15 | 河南大学 | Hyperspectral image classification method of dual-path small convolution network |
| CN113361485A (en) * | 2021-07-08 | 2021-09-07 | 齐齐哈尔大学 | Hyperspectral image classification method based on spectral space attention fusion and deformable convolution residual error network |
| CN113743429A (en) * | 2020-05-28 | 2021-12-03 | 中国人民解放军战略支援部队信息工程大学 | Hyperspectral image classification method and device |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2332270A1 (en) * | 2008-08-19 | 2011-06-15 | Alcatel-Lucent USA Inc. | System and method for receiving high spectral efficiency optical dpsk signals |
| CN102819839A (en) * | 2012-07-19 | 2012-12-12 | 北京市遥感信息研究所 | High-precision registration method for multi-characteristic and multilevel infrared and hyperspectral images |
| CN104063897A (en) * | 2014-06-28 | 2014-09-24 | 南京理工大学 | Satellite hyper-spectral image compressed sensing reconstruction method based on image sparse regularization |
| CN104281855A (en) * | 2014-09-30 | 2015-01-14 | 西安电子科技大学 | Hyperspectral image classification method based on multi-task low rank |
| CN107292343A (en) * | 2017-06-23 | 2017-10-24 | 中南大学 | A kind of Classification of hyperspectral remote sensing image method based on six layers of convolutional neural networks and spectral space information consolidation |
| CN107316013A (en) * | 2017-06-14 | 2017-11-03 | 西安电子科技大学 | Hyperspectral image classification method with DCNN is converted based on NSCT |
-
2018
- 2018-03-23 CN CN201810243423.4A patent/CN108491849B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2332270A1 (en) * | 2008-08-19 | 2011-06-15 | Alcatel-Lucent USA Inc. | System and method for receiving high spectral efficiency optical dpsk signals |
| CN102819839A (en) * | 2012-07-19 | 2012-12-12 | 北京市遥感信息研究所 | High-precision registration method for multi-characteristic and multilevel infrared and hyperspectral images |
| CN104063897A (en) * | 2014-06-28 | 2014-09-24 | 南京理工大学 | Satellite hyper-spectral image compressed sensing reconstruction method based on image sparse regularization |
| CN104281855A (en) * | 2014-09-30 | 2015-01-14 | 西安电子科技大学 | Hyperspectral image classification method based on multi-task low rank |
| CN107316013A (en) * | 2017-06-14 | 2017-11-03 | 西安电子科技大学 | Hyperspectral image classification method with DCNN is converted based on NSCT |
| CN107292343A (en) * | 2017-06-23 | 2017-10-24 | 中南大学 | A kind of Classification of hyperspectral remote sensing image method based on six layers of convolutional neural networks and spectral space information consolidation |
Cited By (25)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109784347A (en) * | 2018-12-17 | 2019-05-21 | 西北工业大学 | Image classification method based on multiple dimensioned dense convolutional neural networks and spectrum attention mechanism |
| CN109784347B (en) * | 2018-12-17 | 2022-04-26 | 西北工业大学 | Image classification method based on multi-scale dense convolution neural network and spectral attention mechanism |
| CN109903255A (en) * | 2019-03-04 | 2019-06-18 | 北京工业大学 | A kind of high spectrum image Super-Resolution method based on 3D convolutional neural networks |
| CN109978041A (en) * | 2019-03-19 | 2019-07-05 | 上海理工大学 | A kind of hyperspectral image classification method based on alternately update convolutional neural networks |
| CN109978041B (en) * | 2019-03-19 | 2022-11-29 | 上海理工大学 | Hyperspectral image classification method based on alternative updating convolutional neural network |
| CN110009015A (en) * | 2019-03-25 | 2019-07-12 | 西北工业大学 | EO-1 hyperion small sample classification method based on lightweight network and semi-supervised clustering |
| CN110263914A (en) * | 2019-05-29 | 2019-09-20 | 朗坤智慧科技股份有限公司 | Deep neural network structure improved method and system based on dense connection |
| CN110852227A (en) * | 2019-11-04 | 2020-02-28 | 中国科学院遥感与数字地球研究所 | Hyperspectral image deep learning classification method, device, equipment and storage medium |
| CN110852369A (en) * | 2019-11-06 | 2020-02-28 | 西北工业大学 | Hyperspectral image classification method combining 3D/2D convolutional network and adaptive spectrum unmixing |
| CN111222453A (en) * | 2020-01-03 | 2020-06-02 | 武汉大学 | Remote sensing image change detection method based on dense connection and geometric structure constraint |
| CN111222453B (en) * | 2020-01-03 | 2022-06-14 | 武汉大学 | Remote sensing image change detection method based on dense connection and geometric structure constraint |
| CN111077093A (en) * | 2020-01-10 | 2020-04-28 | 安徽理工大学 | Method and device for quickly detecting coal gangue based on multispectral technology |
| CN113743429A (en) * | 2020-05-28 | 2021-12-03 | 中国人民解放军战略支援部队信息工程大学 | Hyperspectral image classification method and device |
| CN111797941A (en) * | 2020-07-20 | 2020-10-20 | 中国科学院长春光学精密机械与物理研究所 | Image classification method and system carrying spectral information and spatial information |
| CN112052755A (en) * | 2020-08-24 | 2020-12-08 | 西安电子科技大学 | Semantic convolution hyperspectral image classification method based on multi-path attention mechanism |
| CN111950714A (en) * | 2020-08-24 | 2020-11-17 | 重庆市云迈科技有限公司 | Energy spectrum CT image domain material identification method based on 3D full convolution neural network |
| CN112116563A (en) * | 2020-08-28 | 2020-12-22 | 南京理工大学 | Hyperspectral image target detection method and system based on spectral dimension and space cooperation neighborhood attention |
| CN112183399A (en) * | 2020-09-30 | 2021-01-05 | 珠海欧比特宇航科技股份有限公司 | Hyperspectral image classification method and medium based on separable convolution and dense connection |
| CN112633401A (en) * | 2020-12-29 | 2021-04-09 | 中国科学院长春光学精密机械与物理研究所 | Hyperspectral remote sensing image classification method, device, equipment and storage medium |
| CN112633401B (en) * | 2020-12-29 | 2022-11-08 | 中国科学院长春光学精密机械与物理研究所 | Hyperspectral remote sensing image classification method, device, equipment and storage medium |
| CN112966598A (en) * | 2021-03-04 | 2021-06-15 | 河南大学 | Hyperspectral image classification method of dual-path small convolution network |
| CN112966598B (en) * | 2021-03-04 | 2022-08-30 | 河南大学 | Hyperspectral image classification method of dual-path small convolution network |
| CN112950632A (en) * | 2021-04-18 | 2021-06-11 | 吉林大学 | Coal quality detection method based on hyperspectral imaging technology and convolutional neural network |
| CN112950632B (en) * | 2021-04-18 | 2022-10-14 | 吉林大学 | Coal quality detection method based on hyperspectral imaging technology and convolutional neural network |
| CN113361485A (en) * | 2021-07-08 | 2021-09-07 | 齐齐哈尔大学 | Hyperspectral image classification method based on spectral space attention fusion and deformable convolution residual error network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108491849B (en) | 2021-09-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108491849A (en) | Hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks | |
| CN108182441B (en) | Parallel multichannel convolutional neural network, construction method and image feature extraction method | |
| CN110348399B (en) | Hyperspectral intelligent classification method based on prototype learning mechanism and multidimensional residual error network | |
| Luus et al. | Multiview deep learning for land-use classification | |
| CN106845418A (en) | A kind of hyperspectral image classification method based on deep learning | |
| CN107392925A (en) | Remote sensing image terrain classification method based on super-pixel coding and convolutional neural networks | |
| CN106650830A (en) | Deep model and shallow model decision fusion-based pulmonary nodule CT image automatic classification method | |
| CN112836773A (en) | Hyperspectral image classification method based on global attention residual error network | |
| CN107358260B (en) | Multispectral image classification method based on surface wave CNN | |
| CN108596138A (en) | A kind of face identification method based on migration hierarchical network | |
| CN107679250A (en) | A kind of multitask layered image search method based on depth own coding convolutional neural networks | |
| CN107180248A (en) | Strengthen the hyperspectral image classification method of network based on associated losses | |
| CN105139004A (en) | Face expression identification method based on video sequences | |
| CN103020265B (en) | The method and system of image retrieval | |
| CN107832797B (en) | Multispectral image classification method based on depth fusion residual error network | |
| CN108010034A (en) | Commodity image dividing method and device | |
| Monteiro et al. | Watershed framework to region-based image segmentation | |
| CN110490849A (en) | Surface Defects in Steel Plate classification method and device based on depth convolutional neural networks | |
| CN107424145A (en) | The dividing method of nuclear magnetic resonance image based on three-dimensional full convolutional neural networks | |
| CN107292336A (en) | A kind of Classification of Polarimetric SAR Image method based on DCGAN | |
| CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
| CN110533683A (en) | A kind of image group analysis method merging traditional characteristic and depth characteristic | |
| CN108388904B (en) | Dimensionality reduction method based on convolutional neural network and covariance tensor matrix | |
| CN109344898A (en) | Convolutional neural networks image classification method based on sparse coding pre-training | |
| CN108564111A (en) | A kind of image classification method based on neighborhood rough set feature selecting |
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 |