CN110321941A - The Compression of hyperspectral images and classification method of identifiable feature learning - Google Patents
The Compression of hyperspectral images and classification method of identifiable feature learning Download PDFInfo
- Publication number
- CN110321941A CN110321941A CN201910551561.3A CN201910551561A CN110321941A CN 110321941 A CN110321941 A CN 110321941A CN 201910551561 A CN201910551561 A CN 201910551561A CN 110321941 A CN110321941 A CN 110321941A
- Authority
- CN
- China
- Prior art keywords
- feature
- compression
- sample
- classification
- encoder
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—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
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/42—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals characterised by implementation details or hardware specially adapted for video compression or decompression, e.g. dedicated software implementation
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/44—Decoders specially adapted therefor, e.g. video decoders which are asymmetric with respect to the encoder
Abstract
The invention discloses the Compression of hyperspectral images and classification method of a kind of identifiable feature learning, the technical issues of the practicability is poor for solving existing Compression of hyperspectral images and classification method.Technical solution is the end-to-end compression sorter network comprising two-way branched structure.It is all the way storehouse from the module of coding, to carry out the study of data identifiability feature, wherein encoder is decompressed for carrying out Feature Compression, decoder for feature.All data all pass through encoder and decoder and calculate mean square error loss function;Another way is categorization module, to carry out can diagnostic characteristics classification, encoder and classifier used a shared module, and classifier is classified using the identifiable compressive features that encoder obtains, and completes end-to-end a Feature Compression and classification task.The present invention, which shares coder module, can not only obtain with identifiable feature, and can efficiently carry out classification hyperspectral imagery task according to identifiable feature, and practicability is good.
Description
Technical field
The present invention relates to a kind of Compression of hyperspectral images and classification methods, more particularly to a kind of identifiable feature learning
Compression of hyperspectral images and classification method.
Background technique
Different from traditional images, each pixel of high spectrum image includes a series of continuous spectral bands, is enriched
Spectral information so that it is all applied in numerous areas.Due to spectral band abundant, high spectrum image is compared to conventional coloured silk
Chromatic graph picture consumes more carrying costs in transmission and calculating and calculates cost, especially prominent in satellite remote sensing field.Closely
The method of Nian Lai, deep learning are more and more applied in classification hyperspectral imagery task, but based on full connection or
The model of convolutional neural networks usually requires huge parameter and calculation amount, how while keeping distinctive, reduces model
Parameter amount or high-spectral data amount are current urgent problems to be solved.In order to reduce hyperspectral image data amount, most effective side
Formula is exactly the redundancy removed in data.Common mode includes following several.1) based on the Compression of hyperspectral images of prediction
Method is intended to explore the relevance between pixel or between wave band, document " Mielikainen J, Huang B.Lossless
Compression of Hyperspectral Images Using Clustered Linear Prediction With
Adaptive Prediction Length[J].IEEE Geoscience and Remote Sensing Letters,
2012,9 (6): a fixation is used using Differential Pulse Code Modulation using the correlation between pixel in 1118---1121. "
Length adaptively packed pixel point reduces data volume.2) method of feature conversion utilizes eigentransformation function, will be original
Data are mapped to another feature space to reduce data volume, document " Du Q, Fowler J E.Hyperspectral Image
Compression Using JPEG2000and Principal Component Analysis[J].IEEE
Geoscience&Remote Sensing Letters, 2007,4 (2): using based on wavelet transform in 201-205. "
JPEG compression method carries out the compression of spectral signature.3) Hyperspectral image compression algorithm based on waveband selection selects whole waves
The a subset of section is as compressed data, but inevitably band letter during waveband selection reduces data volume
Breath loss, document " Martinez-Uso A, Pla F,Sotoca,et al.Clustering-Based
Hyperspectral Band Selection Using Information Measures[J].IEEE Transactions
On Geoscience and Remote Sensing, 2008,45 (12): 4158-4171. " proposes a kind of hierarchical clustering side
Case, to reduce the data volume of given high-spectral data, the program uses band grouping method, is minimizing the same of cluster internal variance
When maximize cluster between variance, thus for hyperspectral classification task choosing be suitble to mass center wave band.However above scheme is not examined
Consider the association between classification and compression duty, this compressed data that above method is extracted is applied to comprising deep learning method
It will cause the decline of classification performance when some classifiers inside.Thus how to learn with can distinguishing ability compressive features pair
Classification hyperspectral imagery task is most important.
Summary of the invention
In order to overcome the shortcomings of existing Compression of hyperspectral images and classification method, the practicability is poor, and the present invention provides one kind and can reflect
The Compression of hyperspectral images and classification method of other property feature learning.This method includes that the end-to-end compression of two-way branched structure is classified
Network.It is all the way storehouse from the module of coding, to carry out the study of data identifiability feature, wherein encoder is for carrying out
Feature Compression, decoder are decompressed for feature.All data all pass through encoder and decoder and calculate mean square error loss
Function;Another way is categorization module, to carry out can diagnostic characteristics classification, encoder and classifier used one it is shared
Module, classifier are classified using the identifiable compressive features that encoder obtains, and complete an end-to-end Feature Compression
And classification task.Framework of the present invention by joint a compression and classification, shared coder module not only available tool
There is identifiable feature, high-spectral data is compressed in the identifiable feature space compared with low dimensional, so that decoder
Initial data can be ideally restored, while classifier can efficiently carry out classification hyperspectral imagery according to identifiable feature
Task, practicability are good.
The technical solution adopted by the present invention to solve the technical problems is: a kind of high-spectrum of identifiability feature learning
As compression and classification method, its main feature is that the following steps are included:
Step 1: high-spectral data to be divided into the sample of tape labelWith the sample X=[x of no label1,
x2...,xN]∈RB×N, wherein the tag representation of tape label sample beThe sample of tape label
A batch input is collectively constituted with the sample of no label, the input as entire model.
Step 2: batch sample input coding device is mapped to one with feature identifiability by neural network
Space generates identifiable characteristic Zt。
Step 3: by all identifiable characteristic ZstInput decoder, by the mapping of neural network, after being decompressed
The loss of mean square error is calculated with original input data for high-spectral data.feAnd fdRespectively indicate encoder and decoding
Device.
Step 4: the identifiable feature input categorization module of tape label sample obtains sample by the mapping of neural network
This prediction label and true tag, which calculates, intersects entropy loss.Formula is as follows, wherein θcIndicate model parameter,Indicate prediction
Label.
Step 5: the loss that classifier obtains and the loss that decoder obtains are obtained training by a linear combination
Total losses carries out parameter update using backpropagation and stochastic gradient descent algorithm.
Wherein, lmseIndicate mean square error loss, lceIt indicates to intersect entropy loss, α indicates weight coefficient.
The beneficial effects of the present invention are: this method includes the end-to-end compression sorter network of two-way branched structure.It is all the way
Storehouse is from the module of coding, and to carry out the study of data identifiability feature, wherein encoder is solved for carrying out Feature Compression
Code device is decompressed for feature.All data all pass through encoder and decoder and calculate mean square error loss function;Another way
For categorization module, to carry out can diagnostic characteristics classification, encoder and classifier have used a shared module, classifier
Classified using the identifiable compressive features that encoder obtains, completes end-to-end a Feature Compression and classification task.
The present invention by the framework of joint a compression and classification, shared coder module not only it is available have it is identifiable
High-spectral data is compressed in the identifiable feature space compared with low dimensional by feature, and decoder is ideally gone back
Former initial data, while classifier can efficiently carry out classification hyperspectral imagery task, practicability according to identifiable feature
It is good.
It elaborates With reference to embodiment to the present invention.
Specific embodiment
Define the high-spectral data X=[x comprising N number of pixel1,x2...,xN]∈RB×N, wherein each pixel
Include B wave band, xiIndicate the spectroscopic data of i-th of sample.For the classification problem of a L class, training setAltogether
Comprising M training sample, the tag representation being associated isClassification hyperspectral imagery is appointed
Business is by training set XtWith Yt, it is that each of X pixel assigns a prediction label.
The present invention relates to coder module, decoder module and classifier modules.Coder module is used for tape label number
According to the latent space for being mapped in a holding data identifiability with no label data;Decoder module is responsible for the compression of latent space
Data are restored as far as possible, and here, decoder and encoder have full symmetric structure;Classifier modules are for mapping data
To classification space.Separately below from coder module, decoder module, three aspects of classifier modules and loss function are situated between
It continues
Specific step is as follows for the Compression of hyperspectral images and classification method of identifiability feature learning of the invention:
1, encoder and decoder module.
Initial data passes through encoder, the feature space of a holding data identifiability is mapped the data into, after being used for
Continuous decoder and classifier carry out data decompression and classification.
Encoder is made of one 5 layers of convolutional neural networks, and one batch of normalization layer and modified line are all connect after each layer
Property unit, convolution in encoder using step-length greater than 1 carries out Feature Dimension Reduction, and five layers of convolution kernel size are respectively set as 8,3,
3, d, 1, wherein d represents the 4th layer of characteristic dimension;Convolution step-length is respectively set as 3,2,2,1,1 in five layers, by the 4th
Layer convolution, the dimension of feature become 1, and then the present invention carries out number of features matching using the convolution that size is 1;In encoder institute
In 5 layer networks selected, first 4 layers of convolution kernel number is 512, and the convolution kernel number of layer 5 is the dimension of data after compression
Degree.
Decoder possesses the structure similar with encoder, but changes convolution operation into deconvolution operation in the module, uses
To generate the high spectrum image after decompression, the last layer is matched using the convolution that size is 1 for characteristic dimension.
High-spectral data collection is defined as X, the form that encoder and decoder can be defined as:
Wherein, fe() and fd() respectively indicates coder module and decoder module, θeAnd θdRespectively indicate encoder
With the model parameter of decoder, Z=[z1,z2,...,zn]∈RC×NFor the output of encoder, indicate that the identifiability learnt is special
Sign,The high-spectral data decompressed by Z is indicated, since encoder is only in spectral signature
Space carries out data compression, ziWithRespectively indicate the spectral vector x in XiBy compressed distinctive feature and from ziIn
Decompress the approximate spectral vector rebuild.
2, classifier modules.
Classifier is made of four layers of convolutional neural networks, and one batch of normalization layer is all connect after each layer and amendment is linear single
Member, it is 8,3,3 convolution kernel that three first layers use convolution size respectively, and corresponding convolution kernel number is 16,32,64, convolution step-length
It is disposed as 1, the 4th layer uses a full articulamentum, by normalizing exponential function for compressed distinctive feature ziConversion
To probability distribution space to obtain final prediction label.
For categorization module, prediction label can be defined as form,
Wherein, φ and θcRespectively indicate classifier and its model parameter, ZtIndicate what tape label sample was obtained by encoder
Identifiable feature,Indicate trained label corresponding with marker samples.
3, loss function.
The parameter of neural network is updated according to loss function, therefore it is very heavy to define a suitable loss function
It wants.For identifiable Feature Compression according to the present invention and classification task, the present invention uses mean square error as coding certainly
One loss function uses cross entropy as one loss function of classification, and total loss function form formula is as follows:
Wherein, lmseIndicate mean square error, lceIndicate cross entropy, α is a coefficient of balance, and value is 2 in experiment.With
ytRespectively indicate the label and sample true tag that categorization module is predicted.In the calculating process of loss function, tape label
Sample calculates cross entropy by encoder and classifier modules, and includes tape label and all pass through without all data including label
It crosses encoder and decoder and calculates mean square error loss.
Structure proposed by the present invention is a model end to end, can continuously be led, can be in network parameter learning process
Using stochastic gradient descent method by backpropagation come more new encoder, the parameter of decoder and categorization module, to realize bloom
Modal data can diagnostic characteristics study, so that data compression and image classification task be effectively performed.
The present invention is tested on the high-spectral data collection of three standards, be used only original spectroscopic data without
Under conditions of considering spatial information, ' on India pine tree ' data set, the accuracy rate of classification is 88.77%, and Y-PSNR is
58.98, compared to the method for classical independent component analysis, classification accuracy improves 19.90%, and Y-PSNR improves
13.49;' on university, Pavia ' data set, classification accuracy 91.12%, Y-PSNR 50.36, compared to classics
The compression method of independent component analysis, classification accuracy improve 4.27%, and Y-PSNR improves 1.93;In ' Kennedy
On space center ' data set, classification accuracy 85.06%, Y-PSNR 34.61, compared to classical independent element point
The compression method of analysis, classification accuracy improve 19%, and Y-PSNR improves 2.15.
Claims (1)
1. a kind of Compression of hyperspectral images and classification method of identifiability feature learning, it is characterised in that the following steps are included:
Step 1: high-spectral data to be divided into the sample of tape labelWith the sample X=[x of no label1,x2...,xN]
∈RB×N, wherein the tag representation of tape label sample beThe sample of tape label and without label
Sample collectively constitutes a batch input, the input as entire model;
Step 2: batch sample input coding device is mapped to the space with feature identifiability by neural network,
Generate identifiable characteristic Zt;
Step 3: by all identifiable characteristic ZstInput decoder, the bloom by the mapping of neural network, after being decompressed
The loss of mean square error is calculated with original input data for modal data;feAnd fdRespectively indicate encoder and decoder;
Step 4: the identifiable feature input categorization module of tape label sample obtains sample by the mapping of neural network
Prediction label and true tag, which calculate, intersects entropy loss;Formula is as follows, wherein θcIndicate model parameter,Indicate pre- mark
Label;
Step 5: the loss that classifier obtains and the loss that decoder obtains are obtained total damage of training by a linear combination
It loses, carries out parameter update using backpropagation and stochastic gradient descent algorithm;
Wherein, lmseIndicate mean square error loss, lceIt indicates to intersect entropy loss, α indicates weight coefficient.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910551561.3A CN110321941A (en) | 2019-06-24 | 2019-06-24 | The Compression of hyperspectral images and classification method of identifiable feature learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910551561.3A CN110321941A (en) | 2019-06-24 | 2019-06-24 | The Compression of hyperspectral images and classification method of identifiable feature learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110321941A true CN110321941A (en) | 2019-10-11 |
Family
ID=68120094
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910551561.3A Pending CN110321941A (en) | 2019-06-24 | 2019-06-24 | The Compression of hyperspectral images and classification method of identifiable feature learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110321941A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111770344A (en) * | 2020-07-23 | 2020-10-13 | 陕西理工大学 | Laser spectrum image compression method and system based on deep learning network |
CN112149712A (en) * | 2020-08-19 | 2020-12-29 | 中国地质大学(武汉) | Efficient construction method of hyperspectral remote sensing data compression and classification model |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108491864A (en) * | 2018-02-27 | 2018-09-04 | 西北工业大学 | Based on the classification hyperspectral imagery for automatically determining convolution kernel size convolutional neural networks |
CN108734199A (en) * | 2018-04-24 | 2018-11-02 | 西北工业大学 | High spectrum image robust classification method based on segmentation depth characteristic and low-rank representation |
CN109034224A (en) * | 2018-07-16 | 2018-12-18 | 西安电子科技大学 | Hyperspectral classification method based on double branching networks |
CN109359623A (en) * | 2018-11-13 | 2019-02-19 | 西北工业大学 | High spectrum image based on depth Joint Distribution adaptation network migrates classification method |
CN109543727A (en) * | 2018-11-07 | 2019-03-29 | 复旦大学 | A kind of semi-supervised method for detecting abnormality based on competition reconstruct study |
CN109753973A (en) * | 2018-12-21 | 2019-05-14 | 西北工业大学 | High spectrum image change detecting method based on Weighted Support Vector |
-
2019
- 2019-06-24 CN CN201910551561.3A patent/CN110321941A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108491864A (en) * | 2018-02-27 | 2018-09-04 | 西北工业大学 | Based on the classification hyperspectral imagery for automatically determining convolution kernel size convolutional neural networks |
CN108734199A (en) * | 2018-04-24 | 2018-11-02 | 西北工业大学 | High spectrum image robust classification method based on segmentation depth characteristic and low-rank representation |
CN109034224A (en) * | 2018-07-16 | 2018-12-18 | 西安电子科技大学 | Hyperspectral classification method based on double branching networks |
CN109543727A (en) * | 2018-11-07 | 2019-03-29 | 复旦大学 | A kind of semi-supervised method for detecting abnormality based on competition reconstruct study |
CN109359623A (en) * | 2018-11-13 | 2019-02-19 | 西北工业大学 | High spectrum image based on depth Joint Distribution adaptation network migrates classification method |
CN109753973A (en) * | 2018-12-21 | 2019-05-14 | 西北工业大学 | High spectrum image change detecting method based on Weighted Support Vector |
Non-Patent Citations (3)
Title |
---|
LEI ZHANG 等: "Learning Discriminative Compact Representation for Hyperspectral Imagery Classification", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》 * |
刘冰 等: "面对高光谱影像分类的半监督阶梯网络", 《测绘科学技术学报》 * |
张国东 等: "基于栈式自编码神经网络对高光谱遥感图像分类研究", 《红外技术》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111770344A (en) * | 2020-07-23 | 2020-10-13 | 陕西理工大学 | Laser spectrum image compression method and system based on deep learning network |
CN111770344B (en) * | 2020-07-23 | 2021-10-22 | 陕西理工大学 | Laser spectrum image compression method and system based on deep learning network |
CN112149712A (en) * | 2020-08-19 | 2020-12-29 | 中国地质大学(武汉) | Efficient construction method of hyperspectral remote sensing data compression and classification model |
CN112149712B (en) * | 2020-08-19 | 2023-06-06 | 中国地质大学(武汉) | Efficient hyperspectral remote sensing data compression and classification model construction method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110442684B (en) | Class case recommendation method based on text content | |
CN106227851A (en) | Based on the image search method searched for by depth of seam division that degree of depth convolutional neural networks is end-to-end | |
CN111191514A (en) | Hyperspectral image band selection method based on deep learning | |
CN111814685B (en) | Hyperspectral image classification method based on double-branch convolution self-encoder | |
CN109858557B (en) | Novel semi-supervised classification method for hyperspectral image data | |
Yang et al. | Evolutionary clustering based vector quantization and SPIHT coding for image compression | |
CN111401426B (en) | Small sample hyperspectral image classification method based on pseudo label learning | |
CN108171279A (en) | A kind of adaptive product Grassmann manifold Subspace clustering methods of multi-angle video | |
CN108960333A (en) | Lossless compression method for high spectrum image based on deep learning | |
CN108446582A (en) | Hyperspectral image classification method based on textural characteristics and affine propagation clustering algorithm | |
CN110321941A (en) | The Compression of hyperspectral images and classification method of identifiable feature learning | |
CN107239781A (en) | A kind of super spectral reflectivity method for reconstructing based on RGB image | |
CN110188827A (en) | A kind of scene recognition method based on convolutional neural networks and recurrence autocoder model | |
CN115019104A (en) | Small sample remote sensing image classification method and system based on multi-source domain self-attention | |
CN105719323A (en) | Hyperspectral dimension reducing method based on map optimizing theory | |
CN116312860B (en) | Agricultural product soluble solid matter prediction method based on supervised transfer learning | |
Teimoorinia et al. | Assessment of Astronomical Images Using Combined Machine-Learning Models | |
CN116958709A (en) | Hyperspectral classification method of lightweight tensor convolution long-short-term memory network | |
CN112149712A (en) | Efficient construction method of hyperspectral remote sensing data compression and classification model | |
Singh et al. | A Pre-processing framework for spectral classification of hyperspectral images | |
CN104732246B (en) | A kind of semi-supervised coorinated training hyperspectral image classification method | |
CN109460772B (en) | Spectral band selection method based on information entropy and improved determinant point process | |
CN114998725A (en) | Hyperspectral image classification method based on adaptive spatial spectrum attention kernel generation network | |
CN109614984A (en) | A kind of homologous image detecting method and system | |
CN113628111B (en) | Hyperspectral image super-resolution method based on gradient information constraint |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20191011 |