CN106815601B - Hyperspectral image classification method based on recurrent neural network - Google Patents

Hyperspectral image classification method based on recurrent neural network Download PDF

Info

Publication number
CN106815601B
CN106815601B CN201710014713.7A CN201710014713A CN106815601B CN 106815601 B CN106815601 B CN 106815601B CN 201710014713 A CN201710014713 A CN 201710014713A CN 106815601 B CN106815601 B CN 106815601B
Authority
CN
China
Prior art keywords
sample
feature
high spectrum
recurrent neural
image
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.)
Active
Application number
CN201710014713.7A
Other languages
Chinese (zh)
Other versions
CN106815601A (en
Inventor
张向荣
焦李成
姜凯
安金梁
冯婕
李阳阳
侯彪
马文萍
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian University of Electronic Science and Technology
Original Assignee
Xian University of Electronic Science and Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xian University of Electronic Science and Technology filed Critical Xian University of Electronic Science and Technology
Priority to CN201710014713.7A priority Critical patent/CN106815601B/en
Publication of CN106815601A publication Critical patent/CN106815601A/en
Application granted granted Critical
Publication of CN106815601B publication Critical patent/CN106815601B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • 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
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2136—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on sparsity criteria, e.g. with an overcomplete basis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of hyperspectral image classification methods based on recurrent neural network, it is weaker mainly to solve existing method input feature vector identification, local spatial feature extracts insufficient problem, implementation step includes: the spatial texture feature and rarefaction representation feature of 1. extraction high spectrum images, and to its stacked combination at low-level feature；2. extracting sample local space sequence signature on low-level feature；3. constructing recurrent neural networks model according to local space sequence signature, and utilize training sample local space sequence signature training recurrent neural networks model parameter；4. test sample local space sequence signature is inputted trained recurrent neural networks model, the high-level semantics features of high abstraction are obtained, the classification information of test sample is obtained.The method that the present invention uses deep learning, improves the accuracy of classification hyperspectral imagery, can be used for vegetation investigation, disaster surveillance, and cartography and information obtain.

Description

Hyperspectral image classification method based on recurrent neural network
Technical field
The invention belongs to technical field of image processing, are related to a kind of hyperspectral image classification method, it is distant to can be used for EO-1 hyperion Feel the classification of image.
Background technique
Currently, the spectral resolution with remote sensor is continuously improved, cognition of the people to object spectrum attribute, feature Also it deepens continuously, the characters of ground object that many is hidden within the scope of narrow spectrum is gradually found, has greatly accelerated remote sensing technology Development, so that high-spectrum remote-sensing is become one of most important research direction of 21 century remote sensing fields.
Different from multispectral remote sensing, for spectral remote sensing technology using imaging spectrometer with nanoscale spectral resolution, use is several Ten or several hundred a wave bands are simultaneously imaged earth's surface object, can obtain the continuous spectrum information of atural object, the spy with " collection of illustrative plates " Property, national economy, in terms of all play an important role, have been widely used for landmarks categorizations, target visit The fields such as survey, agricultural monitoring, mineral map plotting, environmental management and national defense construction.
The classification of high spectrum image is an important content of target in hyperspectral remotely sensed image processing and application, and final goal is Unique classification logotype is assigned to each pixel in image.Research and analysis for hyperspectral image data classification task, Main is exactly to tell the characters of ground object in image by high-spectral data, that is, passes through analysis original spectrum or other features letter Breath marks off different earth surface areas to come, such as meadow, farmland, waters, cities and towns, bridge, facilitates people's identification and divides Analyse earth's surface situation.
The features such as high in face of hyperspectral image data dimension, data dependence is strong, redundancy height, local space consistency, Common classification method is mainly started with from the following aspects: (1) rarefaction representation and dictionary learning select a small amount of marked sample As dictionary, gone to indicate other samples with the linear combination of dictionary sample；(2) supporting vector machine SVM classifier and its kernel function Correlating transforms, construct novel kernel function, such as polynomial kernel, adapt to the nonlinear Distribution of high-spectral data；(3) it is semi-supervised with Active Learning promotes the classifying quality under small sample；(4) the new proposition of classifier and the combination of multi-categorizer；(5) feature mentions Take with combined transformation etc..Wherein common feature includes: 1) spectral signature, i.e., high spectrum image itself spectral information characteristics and its Related derived character；2) space characteristics, texture, shape, morphological feature including high spectrum image；3) sky-spectrum signature, i.e., it is empty Between the feature that combines of feature and spectral signature.
These features are started with from the inherent characteristic of high spectrum image, special by simply extracting the low layer that transformation obtains Sign.With the burning hot rise of deep learning in recent years, deep learning model framework how is utilized, sufficiently extracting has more preferable indicate Property high spectrum image high-level characteristic, promoted classification hyperspectral imagery precision, increasingly become domestic and foreign scholars fall over each other research Hot spot.
Deep learning is a kind of Feature Extraction Technology, it is developed from nerual network technique, by way of layering pair Low-level feature carries out high abstraction, to obtain the better representation method of feature.Common deep learning frame mainly has stack Self-encoding encoder SAE, depth confidence network DBN, convolutional neural networks CNN, recurrent neural network RNN, they are widely used in The fields such as natural language processing, computer vision, speech recognition, bioinformatics, and achieve extraordinary effect.
Has scholar at present for stack self-encoding encoder SAE, depth confidence network DBN and convolutional neural networks CNN depth It practises model and is introduced into classification hyperspectral imagery.In " Deep Learning-Based Classification of Hyperspectral Data " in, high-spectral data is carried out PCA dimensionality reduction by Yushi Chen, chooses the pixel in rectangular window It connects into be feature vector as local spatial feature, then connected with original spectrum feature, is inputted as low-level feature Build stack self-encoding encoder SAE model.In " Spectral-Spatial Classification of Hyperspectral Data Based on Deep Belief Network " in, Yushi Chen is special using method building low layer same as above Sign, and input the depth confidence network DBN model built.The classifying quality of this method is general, and accuracy is not high, and There are many deficiencies, such as directlys adopt spectral signature as input feature vector, and the random noise for including is too many, and identification is weaker, nothing Method obtains good classifying quality；And in local spatial feature extraction, simply picks all pixels of neighborhood and be not added Processing, the then pixel wherein to differ greatly with center pixel can seriously affect nicety of grading.
Summary of the invention
It is an object of the invention in view of the above shortcomings of the prior art, propose a kind of bloom based on recurrent neural network Image classification method is composed, to construct the degree of purity more better low-level feature of high-class effect, while being reinforced to picture in local space The exploration of correlation between element, improves the effect of important pixel, reduces the influence of useless pixel, and low-level feature is abstracted as and is sentenced The other higher high-level semantics features of property improve nicety of grading to more fully utilize the characteristic of high spectrum image.
To achieve the above object, technical solution of the present invention includes the following:
(1) a panel height spectrum picture is inputted, which includes K pixel, B EO-1 hyperion spectral coverage, c class atural object, Wherein K=K1×K2, K1Indicate the length of high spectrum image, K2Indicate the width of high spectrum image, each pixel of image is one Sample, each sample indicate that the intrinsic dimensionality of sample is B with a feature vector, and 10% sample is selected in every class atural object This composition training sample set, the sample for being left 90% form test sample collection；
(2) the principal component grayscale image of high spectrum image is filtered using Gabor filter, obtains high spectrum image Spatial texture featureWherein R indicates real number field, and g is spatial texture feature vector dimension；
(3) rarefaction representation coefficient that each pixel in high spectrum image is calculated using the method for rarefaction representation, obtains bloom The rarefaction representation feature of spectrogram pictureWherein m is the dimension of rarefaction representation feature vector；
(4) by the spatial texture feature F of high spectrum image1With rarefaction representation feature F2Stacked combination is at high spectrum image Low-level featureL is the dimension of low-level feature vector, l=g+m；
(5) in high spectrum image low-level feature matrix F, window is constructed centered on each sample, extracts the office of sample Portion's space characteristics block, and utilize the local space sequence signature of the similitude building sample between sample；
(6) recurrent neural networks model is constructed by time step long number of number of samples in window, and inputs training sample Local space sequence signature and corresponding class label repetitive exercise recurrent neural networks model parameter, obtain trained recurrence Neural network model；
(7) the local space sequence signature of test sample is input in trained recurrent neural networks model, is obtained Classification category, completes classification.
The present invention has the advantage that
(1) present invention gets up the spatial texture feature and rarefaction representation feature integration of high spectrum image, and this low layer is special Sign had not only included high spectrum image pixel samples local spatial information, but also included rarefaction representation of the pixel samples about other samples Information, such low-level feature degree of purity and identification are higher, more preferable for classification task effect；
(2) present invention extracts the local space sequence of high spectrum image on the basis of high spectrum image local spatial feature Column feature not only obtains local spatial information, also explores the similarity information in local space between each pixel samples, mention The high effect of important pixel, reduces the influence of useless pixel, improves classifying quality；
(3) present invention utilizes the recurrent neural networks models for being usually used in natural language processing field, by recurrent neural net The temporal characteristics of network combine with high spectrum image local space sequence information, and it is empty can to effectively integrate high spectrum image part Between context relation, be extracted as high-level semantics features for low-level feature is abstract, take full advantage of the characteristic of EO-1 hyperion, improve Classification accuracy rate.
Detailed description of the invention
Fig. 1 is flow diagram of the invention；
Fig. 2 is recurrent neural networks model schematic diagram；
Fig. 3 is the Indian Pines image that present invention emulation uses；
Fig. 4 is the classification results comparison diagram of the present invention and existing method to Indian Pines image.
Specific embodiment
Referring to Fig.1, specific implementation step of the invention includes:
Step 1, high spectrum image is inputted.
Input a width three-dimensional matrice high spectrum image, which includes K pixel samples, B EO-1 hyperion spectral coverage, C class atural object, wherein K=K1×K2, K1Indicate the length of high spectrum image, K2The width for indicating high spectrum image, is selected in every class atural object 10% sample is selected as training sample, is left 90% sample as test sample.
Step 2, the spatial texture feature F of high spectrum image is obtained1。
2a) high spectrum image is converted using Principal Component Analysis, k=10 principal component grayscale image before extracting；
The Gabor filter of 4 directions, 3 kinds of scales 2b) is set, that is, 4 different Gabor kernel function directions and 3 are set A different sinusoidal plane wave wavelength, obtains 12 Gabor filters, the kernel function of each Gabor filter is as follows:
Wherein, x'=xcos θ+ysin θ, y'=-xcos θ+ysin θ, x and y indicates coordinate location information, λ indicate sinusoidal The wavelength of plane wave, θ indicate the direction of Gabor kernel function,Indicate phase deviation, σ indicates the standard deviation of Gaussian envelope, γ table Show space aspect ratio；
2c) using 12 Gabor filters set respectively to each principal component ash in preceding k principal component grayscale image Degree figure carries out Gabor filtering, obtains 12 filtered images of each principal component grayscale image；
2d) together by 12 × k filtered image stacks, the spatial texture feature of high spectrum image is obtainedWherein R indicates real number field, and g=12 × k is spatial texture feature vector length.
Step 3, the rarefaction representation feature F of high spectrum image is obtained2。
3a) sample that every class randomly selects 1% from the c class training sample of high spectrum image constructs son as dictionary atom Dictionary, wherein the sub- dictionary of the i-th class is Indicate the i-th class J-th of dictionary atom, j=1,2 ..., mi, miIndicate the dictionary atom number of the i-th class；
3b) c sub- dictionaries are lined up, an entirety is merged into, obtains total structuring dictionary D=[D1 ... Di ... Dc], D ∈ RB×mIt is a two-dimensional matrix, indicates total structuring dictionary that the sub- dictionary of all classes is constituted, m=m1+…+ mi+…+mcIndicate the sum of the sub- dictionary atom number of all categories；
The rarefaction representation vector that each pixel 3c) is solved using orthogonal matching pursuit algorithm, that is, pass through orthogonal matching pursuit The following formula of algorithm optimization obtains rarefaction representation vector α of each pixel x about structuring dictionary D in high spectrum image:
S.t.D α=x
Wherein | | α | |0Expression takes 0 norm to α；
3d) by the rarefaction representation vector α of samples all in high spectrum image, formed according to raw image data corresponding position One three-dimensional rarefaction representation eigenmatrix
Step 4, by spatial texture featureWith rarefaction representation featureStacked combination is got up, and is obtained To the low-level feature matrix of high spectrum imageWherein l=g+m is low-level feature vector length.
Step 5, the local space sequence signature of sample is constructed.
5a) in high spectrum image low-level feature matrix F, centered on each sample x, size is constructed by side length of w=9 For the window of w × w, the local spatial feature block of x is extracted, then the local spatial feature block of x includes w2=81 samples, wherein often A sample is the low-level feature vector of an a length of l；
The similarity for 5b) calculating each sample and central sample x in local spatial feature block according to Euclidean distance formula is big Small, the smaller similarity of Euclidean distance is bigger, and each sample is lined up from big to small according to similarity, obtains sample x's Local space sequence signature
Step 6, training sample set and its corresponding class label training recurrent neural networks model are utilized.
6a) using number of samples in window as time step long number T, constructs input layer and hidden node number is the recurrence of l Neural network model, wherein T=w × w=81；
The local space sequence signature of training sample and corresponding class label 6b) are input to recurrent neural network mould Each of sample local space sequence signature low-level feature vector is inputted each corresponding time step by type, tool Body method is by t-th of low-level feature x in the local space sequence signature of any one training sample xtInput recurrent neural net T-th of time step of network, by the input x of the time steptWith the hidden layer state s of step-length at previous i.e. the t-1t-1Altogether With the hidden layer state s for constituting the time stept:
st=σ (Uxt+Wst-1),
Wherein U indicates that weight matrix of the input layer to hidden layer, the weight matrix of W expression hidden layer to hidden layer, σ indicate non-linear Activation primitive, σ of the invention select ReLU function, then last output o81By the hidden layer state s of the last one time step81 It determines, it may be assumed that
Wherein V indicate hidden layer to output layer weight matrix,Indicate nonlinear activation function, this exampleSelection Softmax function, then using the parameter in the back-propagation method repetitive exercise recurrent neural networks model for passing through the time, repeatedly Stop after generation 200 times, obtains trained recurrent neural networks model.
Step 7, the local space sequence signature of test sample is input in trained recurrent neural networks model, is obtained To classification category, classification is completed.
Effect of the invention can be further illustrated by following emulation experiment:
1. simulated conditions:
Emulation experiment uses Indian Pines image, which is pushed away in June, 1992 by NASA's NASA jet Unloaded visible light/Infrared Imaging Spectrometer AVIRIS into laboratory is obtained in the Indiana northwestward, as shown in figure 3, image Size is 145 × 145, totally 220 wave bands, removes noise and atmosphere and wave band that waters absorbs is there are also 200 wave bands, totally 16 Class terrestrial object information, as shown in table 1.
Emulation experiment is Intel Core i5-4210, dominant frequency 2.90GHz in CPU, inside saves as 10 system of Windows of 8G It is carried out on system with Python.
16 class data in 1 Indian Pines image of table
Classification Item name Number of samples Classification Item name Number of samples
1 Alfalfa 46 9 Oats 20
2 Corn-notill 1428 10 Soybean-notill 972
3 Corn-mintill 830 11 Soybean-mintill 2455
4 Corn 237 12 Soybean-clean 593
5 Grass-pasture 483 13 Wheat 205
6 Grass-trees 730 14 Woods 1265
7 Grass-pasture-mowed 28 15 Buildings-Grass-Trees-Drives 386
8 Hay-windrowed 478 16 Stone-Steal-Towers 93
2. simulation parameter:
Above emulation experiment is unified select 10% as training sample, and remaining 90% as test sample, SVM punish because Son is set as 491；In SRC method, dictionary directly is constituted with training sample, degree of rarefication is set as 10；In SOMP method, directly use Training sample constitutes dictionary, and window size w is set as 9, and degree of rarefication is set as 30；In the present invention, PCA transformation retains preceding 10 masters The frequency of ingredient, Gabor filter selects { 0.25,0.5,0.75 } 3 kinds of scales, direction selection4 kinds of directions Totally 12 filters, the every 1% building dictionary of class sample random selection of rarefaction representation, degree of rarefication are set as 30, window size w setting It is 9, timing node number T is set as 81.
3. emulation content and result:
Classified with existing three kinds of common methods to high spectrum image Indian Pine using the present invention, commonly Three kinds of methods are respectively: being based on supporting vector machine svm classifier method, based on the classification method of rarefaction representation SRC, be based on partial zones The rarefaction representation SOMP classification method in domain.
Classified to Indian Pines image with the present invention with above-mentioned three kinds of common methods, as a result as shown in figure 4, its Middle Fig. 4 (a) is the classification results figure with SVM method, and Fig. 4 (b) is the classification results figure with rarefaction representation SRC method, Fig. 4 (c) It is the classification results figure with SOMP method, Fig. 4 (d) is classification results figure of the invention.Figure 4, it is seen that compared to normal Three kinds of methods, result figure of the invention are more clear completely, and local space consistency and edge consistency are all than existing side Method will be got well, and nicety of grading is higher.
The present invention and other various methods respectively carry out 10 emulation experiments, take the average value of classification results as final point Class accuracy, including whole accuracy (OA), every class average accuracy (AA) and Kappa coefficient (Kappa), as a result such as 2 institute of table Show.
The classification accuracy rate of table 2 present invention and other methods
Method OA (%) AA (%) Kappa
SVM 81.24 74.06 0.79
SRC 68.53 64.23 0.64
SOMP 95.27 83.48 0.95
The present invention 97.17 95.84 0.97
As seen from Table 2, the present invention and existing SOMP method are due to containing local spatial information, compared to only with single picture The SVM method and SRC method of prime information, classification accuracy rate is obviously higher, and the present invention had both incorporated spatial texture feature With the information of rarefaction representation feature, and the information between local space pixel is sufficiently excavated, and will with recurrent neural networks model It is the better high-level semantics features of the higher representative of identification that low-level feature, which extracts, can obtain higher classification accuracy rate, Whole accuracy, average accuracy, other three kinds of methods of being all better than on Kappa coefficient.
To sum up, the spatial texture feature of high spectrum image and rarefaction representation feature integration are low-level feature by the present invention, and It is extracted local space sequence signature on the basis of local spatial feature, and utilizes the recurrent neural net in deep learning frame Network model classifies to high spectrum image, has not only improved the degree of purity and identification of low-level feature, but also explores high-spectrum Similarity information in the local space of picture between each pixel samples improves the effect of important pixel, reduces useless pixel Influence, while the temporal characteristics of recurrent neural network being combined with high spectrum image local space sequence information, can be with Low-level feature is abstracted and is extracted as high-level semantics features, sufficiently by the context relation for effectively integrating high spectrum image local space The characteristic of EO-1 hyperion is utilized, obtains higher discrimination, there is apparent advantage compared with the existing methods.

Claims (5)

1. a kind of hyperspectral image classification method based on recurrent neural network, comprising:
(1) a panel height spectrum picture is inputted, which includes K pixel, B EO-1 hyperion spectral coverage, c class atural object, wherein K =K1×K2, K1Indicate the length of high spectrum image, K2Indicating the width of high spectrum image, each pixel of image is a sample, Each sample indicates that the intrinsic dimensionality of sample is B with a feature vector, and 10% sample composition is selected in every class atural object Training sample set, the sample for being left 90% form test sample collection；
(2) the principal component grayscale image of high spectrum image is filtered using Gabor filter, obtains the space of high spectrum image Textural characteristicsWherein R indicates real number field, and g is spatial texture feature vector dimension；
(3) rarefaction representation coefficient that each pixel in high spectrum image is calculated using the method for rarefaction representation, obtains high-spectrum The rarefaction representation feature of pictureWherein m is the dimension of rarefaction representation feature vector；
(4) by the spatial texture feature F of high spectrum image1With rarefaction representation feature F2Stacked combination at high spectrum image low layer FeatureL is the dimension of low-level feature vector, l=g+m；
(5) in high spectrum image low-level feature matrix F, window is constructed centered on each sample, the part for extracting sample is empty Between characteristic block, and utilize the local space sequence signature of the similitude building sample between sample；
(6) recurrent neural networks model is constructed by time step long number of number of samples in window, and inputs the part of training sample Spatial sequence feature and corresponding class label repetitive exercise recurrent neural networks model parameter, obtain trained recurrent neural Network model；
(7) the local space sequence signature of test sample is input in trained recurrent neural networks model, is classified Category completes classification.
2. according to the method described in claim 1, wherein using Gabor filter to the principal component of high spectrum image in step 2 Grayscale image is filtered, and is carried out as follows:
2a) high spectrum image is converted using Principal Component Analysis, k=10 principal component grayscale image before extracting；
4 different Gabor kernel function directions and 3 different sinusoidal plane wave wavelength 2b) are set, 12 Gabor filters are obtained The kernel function of wave device, each Gabor filter is as follows:
Wherein, x'=xcos θ+ysin θ, y'=-xcos θ+ysin θ, x and y indicates coordinate location information, λ indicate sinusoidal plane The wavelength of wave, θ indicate the direction of Gabor kernel function,Indicate phase deviation, σ indicates the standard deviation of Gaussian envelope, and γ indicates empty Between aspect ratio；
2c) using 12 Gabor filters setting respectively to each principal component grayscale image in k principal component grayscale image into Row Gabor filtering, obtains 12 filtered images of each principal component grayscale image；
2d) together by 12 × k filtered image stacks, the spatial texture feature of high spectrum image is obtainedG=12 × k representation space texture feature vector dimension.
3. according to the method described in claim 1, wherein being calculated in step 3 using the method for rarefaction representation every in high spectrum image The rarefaction representation coefficient of a pixel carries out as follows:
3a) sample that every class randomly selects 1% from the c class training sample of high spectrum image constructs sub- word as dictionary atom Allusion quotation, wherein the sub- dictionary of the i-th class is Indicate the jth of the i-th class A dictionary atom, j=1,2 ..., mi, miIndicate the dictionary atom number of the i-th class；
3b) c sub- dictionaries are lined up, an entirety is merged into, obtains total structuring dictionary D=[D1...Di...Dc], D∈RB×mIt is a two-dimensional matrix, indicates total structuring dictionary that the sub- dictionary of all classes is constituted, m=m1+…+mi+…+mc It is the dimension of rarefaction representation feature vector, it is added by the sub- dictionary atom number of all categories and is obtained；
The rarefaction representation vector that each pixel 3c) is solved using orthogonal matching pursuit algorithm, that is, pass through orthogonal matching pursuit algorithm Optimize following formula, obtain rarefaction representation vector α of each pixel x about structuring dictionary D in high spectrum image:
S.t.D α=x
Wherein | | α | |0Expression takes 0 norm to α；
3d) by the rarefaction representation vector α of samples all in high spectrum image, one is formed according to raw image data corresponding position Three-dimensional rarefaction representation eigenmatrix
4. according to the method described in claim 1, the local space sequence signature of sample is wherein constructed in step 5, by following step It is rapid to carry out:
5a) in high spectrum image low-level feature matrix F, centered on any one sample x, constructed by window side length of w=9 Size is w × w rectangular window, extracts the local spatial feature block of x, i.e., size is w × w × l three-dimensional matrice, then x Local spatial feature block includes w2=81 pixel samples, wherein each sample is the low-level feature vector of an a length of l, l=g + m, g representation space texture feature vector dimension, m indicate rarefaction representation feature vector dimension；
The similarity size of each pixel samples and center pixel sample x in local spatial feature block 5b) is calculated, and by each picture Element is lined up from big to small according to similarity, obtains the local space sequence signature of pixel samples x
5. according to the method described in claim 1, wherein construct recurrent neural networks model and training pattern parameter in step 6, It carries out as follows:
6a) building time step be T, input layer and hidden node number be l recurrent neural networks model, wherein T=w × W=81, w indicate window side length, l=g+m, g representation space texture feature vector dimension, m expression rarefaction representation feature vector dimension Number；
The local space sequence signature of training sample 6b) is input to recurrent neural networks model, i.e., by sample local space sequence Each of column feature low-level feature vector inputs each corresponding time step, and using the backpropagation by the time Parameter in method repetitive exercise recurrent neural networks model, obtains trained recurrent neural networks model.
CN201710014713.7A 2017-01-10 2017-01-10 Hyperspectral image classification method based on recurrent neural network Active CN106815601B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710014713.7A CN106815601B (en) 2017-01-10 2017-01-10 Hyperspectral image classification method based on recurrent neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710014713.7A CN106815601B (en) 2017-01-10 2017-01-10 Hyperspectral image classification method based on recurrent neural network

Publications (2)

Publication Number Publication Date
CN106815601A CN106815601A (en) 2017-06-09
CN106815601B true CN106815601B (en) 2019-10-11

Family

ID=59110109

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710014713.7A Active CN106815601B (en) 2017-01-10 2017-01-10 Hyperspectral image classification method based on recurrent neural network

Country Status (1)

Country Link
CN (1) CN106815601B (en)

Families Citing this family (30)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107194437B (en) * 2017-06-22 2020-04-07 重庆大学 Image classification method based on Gist feature extraction and concept machine recurrent neural network
CN107169535B (en) * 2017-07-06 2023-11-03 谈宜勇 Deep learning classification method and device for biological multispectral image
CN107844751B (en) * 2017-10-19 2021-08-27 陕西师范大学 Method for classifying hyperspectral remote sensing images of guide filtering long and short memory neural network
CN107798348B (en) * 2017-10-27 2020-02-18 广东省智能制造研究所 Hyperspectral image classification method based on neighborhood information deep learning
CN110399929B (en) * 2017-11-01 2023-04-28 腾讯科技（深圳）有限公司 Fundus image classification method, fundus image classification apparatus, and computer-readable storage medium
CN107679525B (en) * 2017-11-01 2022-11-29 腾讯科技（深圳）有限公司 Image classification method and device and computer readable storage medium
CN108171270B (en) * 2018-01-05 2021-08-27 大连海事大学 Hyperspectral image classification method based on Hash learning
CN108256454B (en) * 2018-01-08 2020-08-14 浙江大华技术股份有限公司 Training method based on CNN model, and face posture estimation method and device
CN108460342B (en) * 2018-02-05 2021-01-01 西安电子科技大学 Hyperspectral image classification method based on convolutional neural network and cyclic neural network
CN108764303A (en) * 2018-05-10 2018-11-06 电子科技大学 A kind of remote sensing images spatial term method based on attention mechanism
US10643092B2 (en) 2018-06-21 2020-05-05 International Business Machines Corporation Segmenting irregular shapes in images using deep region growing with an image pyramid
US10776923B2 (en) 2018-06-21 2020-09-15 International Business Machines Corporation Segmenting irregular shapes in images using deep region growing
CN109002771B (en) * 2018-06-26 2022-04-08 中国科学院遥感与数字地球研究所 Remote sensing image classification method based on recurrent neural network
CN109460471B (en) * 2018-11-01 2021-09-24 信融源大数据科技(北京)有限公司 Method for establishing fiber category map library based on self-learning mode
CN109670042A (en) * 2018-12-04 2019-04-23 广东宜教通教育有限公司 A kind of examination question classification and grade of difficulty method based on recurrent neural network
CN109615008B (en) * 2018-12-11 2022-05-13 华中师范大学 Hyperspectral image classification method and system based on stack width learning
CN109711466B (en) * 2018-12-26 2023-04-14 陕西师范大学 CNN hyperspectral image classification method based on edge preserving filtering
CN109816002B (en) * 2019-01-11 2022-09-06 广东工业大学 Single sparse self-encoder weak and small target detection method based on feature self-migration
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
CN110188794B (en) * 2019-04-23 2023-02-28 深圳大学 Deep learning model training method, device, equipment and storage medium
CN110163293A (en) * 2019-05-28 2019-08-23 武汉轻工大学 Red meat classification method, device, equipment and storage medium based on deep learning
CN110363078B (en) * 2019-06-05 2023-08-04 广东三姆森科技股份有限公司 Method and device for classifying hyperspectral images based on ADMM-Net
CN110866439B (en) * 2019-09-25 2023-07-28 南京航空航天大学 Hyperspectral image joint classification method based on multi-feature learning and super-pixel kernel sparse representation
CN110852451B (en) * 2019-11-27 2022-03-01 电子科技大学 Recursive kernel self-adaptive filtering method based on kernel function
CN111582330A (en) * 2020-04-22 2020-08-25 北方民族大学 Integrated ResNet-NRC method for dividing sample space based on lung tumor image
CN111860654B (en) * 2020-07-22 2024-02-02 河南大学 Hyperspectral image classification method based on cyclic neural network
CN113128669A (en) * 2021-04-08 2021-07-16 中国科学院计算技术研究所 Neural network model for semi-supervised learning and semi-supervised learning method
CN113139532B (en) * 2021-06-22 2021-09-21 中国地质大学（武汉） Classification method based on multi-output classification model, computer equipment and medium
CN113887656B (en) * 2021-10-21 2024-04-05 江南大学 Hyperspectral image classification method combining deep learning and sparse representation
CN117649943B (en) * 2024-01-30 2024-04-30 吉林大学 Shaping data intelligent analysis system and method based on machine learning

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103514456A (en) * 2013-06-30 2014-01-15 安科智慧城市技术（中国）有限公司 Image classification method and device based on compressed sensing multi-core learning
CN104036289A (en) * 2014-06-05 2014-09-10 哈尔滨工程大学 Hyperspectral image classification method based on spatial and spectral features and sparse representation
CN104091151A (en) * 2014-06-30 2014-10-08 南京信息工程大学 Vehicle identification method based on Gabor feature extraction and sparse representation
CN104298999A (en) * 2014-09-30 2015-01-21 西安电子科技大学 Hyperspectral feature leaning method based on recursion automatic coding
US9152881B2 (en) * 2012-09-13 2015-10-06 Los Alamos National Security, Llc Image fusion using sparse overcomplete feature dictionaries

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB2532075A (en) * 2014-11-10 2016-05-11 Lego As System and method for toy recognition and detection based on convolutional neural networks

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9152881B2 (en) * 2012-09-13 2015-10-06 Los Alamos National Security, Llc Image fusion using sparse overcomplete feature dictionaries
CN103514456A (en) * 2013-06-30 2014-01-15 安科智慧城市技术（中国）有限公司 Image classification method and device based on compressed sensing multi-core learning
CN104036289A (en) * 2014-06-05 2014-09-10 哈尔滨工程大学 Hyperspectral image classification method based on spatial and spectral features and sparse representation
CN104091151A (en) * 2014-06-30 2014-10-08 南京信息工程大学 Vehicle identification method based on Gabor feature extraction and sparse representation
CN104298999A (en) * 2014-09-30 2015-01-21 西安电子科技大学 Hyperspectral feature leaning method based on recursion automatic coding

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Preprocessing-free surface material classification using convolutional neural networks pretrained by sparse Autoencoder;Mengqi Ji etal.;《2015 IEEE 25th International Workshop on Machine Learning for Signal Processing (MLSP)》;20150920;第1-6页 *
基于稀疏表示的人脸表情识别;朱可;《中国优秀硕士学位论文全文数据库 信息科技辑》;20131215;第9-43页 *

Also Published As

Publication number Publication date
CN106815601A (en) 2017-06-09

Similar Documents

Publication Publication Date Title
CN106815601B (en) Hyperspectral image classification method based on recurrent neural network
Wang et al. Auto-AD: Autonomous hyperspectral anomaly detection network based on fully convolutional autoencoder
Wu et al. ORSIm detector: A novel object detection framework in optical remote sensing imagery using spatial-frequency channel features
Zhao et al. Object-based convolutional neural network for high-resolution imagery classification
Shen et al. Efficient deep learning of nonlocal features for hyperspectral image classification
Xu et al. A lightweight and robust lie group-convolutional neural networks joint representation for remote sensing scene classification
Zhang et al. Weakly supervised learning based on coupled convolutional neural networks for aircraft detection
CN106529508B (en) Based on local and non local multiple features semanteme hyperspectral image classification method
Liu et al. Stacked Fisher autoencoder for SAR change detection
CN103971123B (en) Hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL)
CN106023065B (en) A kind of tensor type high spectrum image spectral-spatial dimension reduction method based on depth convolutional neural networks
CN109389080A (en) Hyperspectral image classification method based on semi-supervised WGAN-GP
Huang et al. Multi-scale local context embedding for LiDAR point cloud classification
CN107247930A (en) SAR image object detection method based on CNN and Selective Attention Mechanism
CN102208034A (en) Semi-supervised dimension reduction-based hyper-spectral image classification method
CN108537121A (en) Self-adaptive remote sensing scene classification method based on meteorological environment parameter and image information fusion
CN107767416A (en) The recognition methods of pedestrian's direction in a kind of low-resolution image
CN109034213B (en) Hyperspectral image classification method and system based on correlation entropy principle
CN108427913A (en) The Hyperspectral Image Classification method of combined spectral, space and hierarchy information
CN108596195A (en) A kind of scene recognition method based on sparse coding feature extraction
Deng A survey of convolutional neural networks for image classification: Models and datasets
CN105160351A (en) Semi-monitoring high-spectral classification method based on anchor point sparse graph
CN114358211B (en) Multi-mode deep learning-based aircraft behavior intention recognition method
Tun et al. Hyperspectral remote sensing images classification using fully convolutional neural network
Xu et al. UCDFormer: Unsupervised change detection using a transformer-driven image translation

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