CN109727210B - Based on approximation L0Remote sensing image unmixing method and system of reconstructed deep belief network - Google Patents
Based on approximation L0Remote sensing image unmixing method and system of reconstructed deep belief networkDownload PDFInfo
- Publication number
- CN109727210B CN109727210BCN201811564932.3ACN201811564932ACN109727210BCN 109727210 BCN109727210 BCN 109727210BCN 201811564932 ACN201811564932 ACN 201811564932ACN 109727210 BCN109727210 BCN 109727210B
- Authority
- CN
- China
- Prior art keywords
- network
- unmixing
- data
- image
- layer
- 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.)
- Expired - Fee Related
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
Description
Technical Field
The invention relates to the field of images, in particular to a method based on approximate L0A remote sensing image unmixing method and system of a modified deep belief network are provided.
Background
The remote sensing image is a film or a photo for recording the electromagnetic wave size of various ground objects and is mainly divided into an aerial photo and a satellite photo; the remote sensing image has larger information amount and higher resolution, so that the classification capability of the remote sensing image on the ground features is greatly increased to a certain extent compared with the prior art.
Most of the conventional image unmixing modes utilize a linear unmixing model, and in order to improve the precision of an abundance matrix after unmixing, an image L is often required to be added to the abundance matrix1Although the accuracy of the unmixed image is relatively improved under a plurality of constraint conditions, the adjustment of a plurality of parameters is troublesome, and the unmixed effect is sensitive to the parameters. In recent years, with the rise of artificial intelligence, deep learning is taken as a research hotspot in the field of artificial intelligence, namely, the network is trained by using the existing sample data from bottom to top, the features of each layer are sequentially extracted, the features are stored by using weights between layers, and after a new sample is input, the trained network can be identified by using the stored weights. The deep belief network is used as an algorithm with superior image processing effect in deep learning, and the unmixing precision is improved while the unmixing mode is maximizedThe quantity is simple and convenient.
Disclosure of Invention
Aiming at the traditional image unmixing method, the technical problem to be solved by the invention is to improve the unmixing precision of the image under the condition of not manually adjusting parameters. The technical scheme is as follows: construction based on approximation L0The remote sensing image unmixing method and system of the improved deep belief network can effectively improve the unmixing precision of the image; the method specifically comprises the following steps:
s1, constructing a group of data by using a spectrum library with a wave band of M and a ground feature type of N, and dividing the data into training data and testing data according to a proportion; wherein M, N are all constants greater than 0;
s2, inputting the training data into a deep belief network for network training;
s3, inputting the test data into the deep belief network trained in the step S2 to obtain a demixing abundance matrix of the test data; judging the image unmixing effect of the current network by using the unmixing abundance matrix of the test data; judging the unmixing accuracy RMSE of the current network image by adopting the quantitative evaluation index, wherein the obtained unmixing accuracy RMSE is larger, namely the marrying effect of the image is poorer, and returning to the step S2 to adjust and change the deep belief network structure and retrain the network; otherwise, the next step S4 is executed;
s4, inputting real remote sensing image data to the depth belief network trained in the step S2 to obtain a unmixing abundance matrix y (k) of the real remote sensing image; the real data used in the method is a spectrum library with G wave bands and H ground object types, and G and H are constants larger than 0.
Further, the deep belief network used in step S2 is formed by stacking 4 restricted boltzmann machines, and the training process is trained from low to high layer by layer, wherein an output layer of each restricted boltzmann machine is used as an input layer of a next restricted boltzmann machine, and the visual layer and the hidden layer of the first 3 restricted boltzmann machines are connected in a directed manner, and the visual layer and the hidden layer of the 4 th restricted boltzmann machine are connected in an undirected manner.
Further, step (ii)In step S2, the parameters of the deep belief network are adjusted by constructing an objective function, and adding an approximate L to the result of the network by minimizing the square error between the output result and the expected result0The norm is used for sparsely solving the mixed abundance, a BP algorithm is used for fine tuning the network weight matrix, so that the output value of the trained network is as close as possible to the expected output value, and the objective function is as follows:
min[e(k)-d(k)]2+λf(α,e(k))，
where d (k) is the actual abundance matrix, e (k) is the unmixed abundance matrix of the net output, L0Norm f (alpha, e (k)) 1/logα(α × e (k)), α is a constant, α → 0 and α ≠ 0; λ is the regularization factor.
Further, in step S3, the network unmixing effect is judged by a quantitative evaluation index, wherein the quantitative evaluation index reflects the accuracy of unmixing through the following expression:
wherein X and
respectively representing a real abundance matrix and an abundance matrix output by a network, wherein n is a wave band value of the simulation data spectrum library, and p is a surface feature type value of the simulation data spectrum library; the smaller the obtained RMSE value is, the better the network image unmixing effect is.
The invention provides a method based on approximate L0The remote sensing image unmixing system of the improved deep belief network specifically comprises the following modules:
the data construction module constructs a group of data through a known spectrum library, and divides the data into training data and test data according to a proportion;
the network training module is used for inputting the training data into a deep belief network for network training; wherein the parameters of the deep belief network are adjusted by constructing an objective function to enable the output of the networkThe square error between the result and the expected result is minimized, and an approximate L is added on the basis of the square error0The norm is used for sparsely solving the mixed abundance, a BP algorithm is used for fine tuning the network weight matrix, so that the output value of the trained network is as close as possible to the expected output value, and the objective function is as follows:
min[e(k)-d(k)]2+λf(α,e(k))，
where d (k) is the actual abundance matrix, e (k) is the unmixed abundance matrix of the net output, L0Norm f (alpha, e (k)) 1/logα(α × e (k)), α is a constant, α → 0 and α ≠ 0; λ is a regularization factor;
the used deep belief network is formed by stacking 4 limited Boltzmann machines, the training process is trained from low to high layer by layer, wherein the output layer of each limited Boltzmann machine is used as the input layer of the next limited Boltzmann machine, the visual layers and the hidden layers of the first 3 limited Boltzmann machines are in directed connection, and the visual layer and the hidden layer of the 4 th limited Boltzmann machine are in undirected connection;
the test network unmixing effect module is used for inputting the test data into the deep belief network trained in the network training module to obtain an unmixing abundance matrix of the test data; judging the image unmixing effect of the current network by using the unmixing abundance matrix of the test data; judging the unmixing accuracy RMSE of the current network image by adopting the quantitative evaluation index, wherein the obtained unmixing accuracy RMSE is larger, namely the marrying effect of the image is poorer, and the depth belief network structure in the network training module is adjusted and the network is retrained at present;
the output module is used for inputting real remote sensing image data to the trained depth belief network in the network training module to obtain a demixing abundance matrix y (k) of the real remote sensing image; the used real remote sensing image data is a spectral library with G wave bands and H ground object types, and G and H are constants larger than 0.
The deep belief network is used as an algorithm with superior image processing effect applied to deep learning, and is applied to remote sensing image solution in the remote sensing image solution mixing method and system provided by the inventionMixing, adding approximate L without artificial parameter adjustment0The norm is a constraint condition, so that the unmixing precision can be further improved.
Based on approximate L in the invention0The invention discloses a remote sensing image unmixing method and system of a modified deep belief network, which are different from the traditional remote sensing image unmixing identification method0The norm improves the depth belief network, and the improved network is applied to remote sensing image unmixing, so that the unmixing mode is simple and quick as possible while the image unmixing precision is improved.
Drawings
The invention will be further described with reference to the accompanying drawings and examples, in which:
FIG. 1 is a deep belief network flow diagram;
FIG. 2 is a block diagram of a deep belief network unmixing system;
FIG. 3 illustrates the use of conventional L's respectively1Norm, DBN network, and DBN network plus approximate L0And (4) performing unmixing on the image by using three norms.
Detailed Description
For a more clear understanding of the technical features, objects and effects of the present invention, embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
Referring to fig. 1, it is a flowchart of an implementation of the method for sparsely unmixing remote sensing images provided by the present invention, and specifically includes the following steps:
l1, constructing a group of data by using the wave band 224 and the spectral library with the ground feature type of 10, and dividing the data into training data and testing data according to a proportion;
l2, inputting the training data into the deep belief network for network training; the network training comprises the following specific steps:
a1, since each feature point may not contain all feature types, the output unmixed abundance matrix e (k) is sparse in nature; based on the sparse characteristics, adding an approximate L on the basis of the unmixing abundance matrix e (k)0Constraint of normConstructing a network cost function under the condition;
a2, using BP algorithm to fine tune weight matrix, making the output value of the network obtained by training as close to the expected output value as possible, but changing the network cost function in the process, constructing an objective function for making the result sparse, and adding approximate L on the basis of minimizing the square error0Norm, and the objective function is specifically:
min[e(k)-d(k)]2+λf(α,e(k))，
wherein the function f (α, e (k)) is 1/log (α, e (k)))α(α × e (k)), α is a constant, α → 0 and α ≠ 0; λ is the regularization factor.
L3, inputting the test data into the deep belief network trained in the step L2 to obtain an unmixed abundance matrix of the test data; judging the image unmixing effect of the current network by using the unmixing abundance matrix of the test data; judging the unmixing accuracy RMSE of the current network image by adopting the quantitative evaluation index, wherein the obtained unmixing accuracy RMSE is larger, namely the marrying effect of the image is poorer, and returning to the step L2 to adjust and change the deep belief network structure and retrain the network; otherwise, the next step L4 is executed;
the image unmixing effect of the current network is judged by using the root mean square error as a quantitative evaluation index. The specific expression is as follows:
wherein X and
respectively representing a real abundance matrix and an abundance matrix output by a network, wherein n is a wave band value of the simulation data spectrum library, and p is a surface feature type value of the simulation data spectrum library; the smaller the obtained RMSE value is, the better the network image unmixing effect is.
L4, inputting real remote sensing image data to the depth belief network trained in the step S2 to obtain a unmixing abundance matrix y (k) of the real remote sensing image; the real data used is a spectrum library with 22 wave bands and 5 ground object types;
please refer to fig. 2, which is a module diagram of a deep belief network unmixing system, and specifically includes adata construction module 21, anetwork training module 22, a testing networkunmixing effect module 23, and anoutput module 24, wherein:
thedata construction module 21 is used for constructing a group of data through different types of spectral libraries and dividing the data into training data and testing data according to a proportion;
thenetwork training module 22 is used for inputting the training data into the deep belief network for network training; wherein, the parameters of the deep belief network are adjusted by constructing an objective function, and the approximate L is added on the basis of minimizing the square error between the output result and the expected result of the network0The norm is used for sparsely solving the mixed abundance, a BP algorithm is used for fine tuning the network weight matrix, so that the output value of the trained network is as close as possible to the expected output value, and the objective function is as follows:
min[e(k)-d(k)]2+λf(α,e(k))，
where d (k) is the actual abundance matrix, e (k) is the unmixed abundance matrix of the net output, L0Norm f (alpha, e (k)) 1/logα(α × e (k)), α is a constant, α → 0 and α ≠ 0; λ is a regularization factor;
the used deep belief network is formed by stacking 4 limited Boltzmann machines, the training process is trained from low to high layer by layer, wherein the output layer of each limited Boltzmann machine is used as the input layer of the next limited Boltzmann machine, the visual layers and the hidden layers of the first 3 limited Boltzmann machines are in directed connection, and the visual layer and the hidden layer of the 4 th limited Boltzmann machine are in undirected connection;
the test networkunmixing effect module 23 is configured to input the test data into a deep belief network trained in the network training module to obtain an unmixing abundance matrix of the test data; judging the image unmixing effect of the current network by using the unmixing abundance matrix of the test data; judging the unmixing accuracy RMSE of the current network image by adopting the quantitative evaluation index, wherein the obtained unmixing accuracy RMSE is larger, namely the marrying effect of the image is poorer, and the depth belief network structure in the network training module is adjusted and the network is retrained at present;
theoutput module 24 is an output module and is used for inputting the real remote sensing image data to the trained depth belief network in the network training module to obtain the unmixed abundance matrix y (k) of the real remote sensing image. The used real remote sensing image data is a spectral library with G wave bands and H ground object types, and G and H are constants larger than 0.
In this embodiment, the DBN network proposed by the present invention is added with approximate L0After comparing the norm with two conventional image unmixing methods, RMSE of the method in 3 is calculated, the number of unified iterations is 20000, and the result is shown in table 1:
TABLE 1 accuracy of each algorithm for real data
| Algorithm | RMSE (root mean square error) | Number of iterations |
| Legacy L1Norm of | 0.1535 | 20000 |
| DBN network | 0.1658 | 20000 |
| DBN network plus approximate L0Norm of | 0.1485 | 20000 |
As can be seen from table 1, the improved DBN network has an improved unmixing accuracy by about two percent compared to the DBN network before the improvement.
Please refer to fig. 3, which respectively use the conventional L1Norm, DBN network, and DBN network plus approximate L0The effect diagram of the image unmixing by the norm three algorithms is shown, and the improved DBN network can be obviously seen from the circled parts of the images before and after the contrast improvement in the diagram, namely, the output abundance is increased by approximate L0After the norm is restrained, the effect of image unmixing by using the norm is better than that of image unmixing by only using a DBN network, and the unmixing precision is effectively improved on the basis of the prior art.
While the present invention has been described with reference to the embodiments shown in the drawings, the present invention is not limited to the embodiments, which are illustrative and not restrictive, and it will be apparent to those skilled in the art that various changes and modifications can be made therein without departing from the spirit and scope of the invention as defined in the appended claims.
Claims (3)
1. A remote sensing image unmixing method based on a depth belief network reconstructed on the basis of approximate L0 is characterized by comprising the following steps:
s1, constructing a group of data by using a spectrum library with a wave band of M and a ground feature type of N, and dividing the data into training data and testing data according to a proportion; wherein M, N are all constants greater than 0;
s2, inputting the training data into a deep belief network for network training; in step S2, the average error between the output result and the expected result of the network is minimized by constructing an objective function to adjust the parameters of the deep belief networkAdding approximate L on the basis0The norm is used for sparsely solving the mixed abundance, a BP algorithm is used for fine tuning the network weight matrix, so that the output value of the trained network is as close as possible to the expected output value, and the objective function is as follows:
min[e(k)-d(k)]2+λf(α,e(k))，
where d (k) is the actual abundance matrix, e (k) is the unmixed abundance matrix of the net output, L0Norm of
α is a constant, α → 0 and α ≠ 0; λ is a regularization factor;
the used deep belief network is formed by stacking 4 limited Boltzmann machines, the training process is trained from low to high layer by layer, wherein the output layer of each limited Boltzmann machine is used as the input layer of the next limited Boltzmann machine, the visual layers and the hidden layers of the first 3 limited Boltzmann machines are in directed connection, and the visual layer and the hidden layer of the 4 th limited Boltzmann machine are in undirected connection;
s3, inputting the test data into the deep belief network trained in the step S2 to obtain a demixing abundance matrix of the test data; judging the image unmixing effect of the current network by using the unmixing abundance matrix of the test data; calculating the image unmixing accuracy RMSE of the current network by adopting the quantitative evaluation index, judging that the image unmixing effect of the current network is poor when the obtained image unmixing accuracy RMSE is greater than or equal to a preset unmixing threshold value, and returning to the step S2 to adjust and change the depth belief network structure and retrain the network again; otherwise, the next step S4 is executed;
s4, inputting real remote sensing image data to the depth belief network trained in the step S2 to obtain a unmixing abundance matrix y (k) of the real remote sensing image; the real data used in the method is a spectrum library with G wave bands and H ground object types, and G and H are constants larger than 0.
2. The remote sensing image unmixing method according to claim 1, wherein the quantitative evaluation index in step S3 is an image unmixing accuracy RMSE of the network calculated by the following expression:
wherein X and
and respectively representing the real abundance matrix and the abundance matrix output by the network, wherein n is a wave band value of the simulation data spectrum library, and p is a ground object type value of the simulation data spectrum library.
3. A remote sensing image unmixing system based on a depth belief network reconstructed on the basis of approximate L0 is characterized by comprising the following modules:
the data construction module constructs a group of data through a known spectrum library, and divides the data into training data and test data according to a proportion;
the network training module is used for inputting the training data into a deep belief network for network training; wherein, the average error between the output result and the expected result of the network is minimized by constructing an objective function to adjust the parameters of the deep belief network, and an approximate L is added on the basis0The norm is used for sparsely solving the mixed abundance, a BP algorithm is used for fine tuning the network weight matrix, so that the output value of the trained network is as close as possible to the expected output value, and the objective function is as follows:
min[e(k)-d(k)]2+λf(α,e(k))，
where d (k) is the actual abundance matrix, e (k) is the unmixed abundance matrix of the net output, L0Norm of
α is a constant, α → 0 and α ≠ 0; λ is a regularization factor;
the used deep belief network is formed by stacking 4 limited Boltzmann machines, the training process is trained from low to high layer by layer, wherein the output layer of each limited Boltzmann machine is used as the input layer of the next limited Boltzmann machine, the visual layers and the hidden layers of the first 3 limited Boltzmann machines are in directed connection, and the visual layer and the hidden layer of the 4 th limited Boltzmann machine are in undirected connection;
the test network unmixing effect module is used for inputting the test data into the deep belief network trained in the network training module to obtain an unmixing abundance matrix of the test data; judging the image unmixing effect of the current network by using the unmixing abundance matrix of the test data; the image unmixing accuracy RMSE of the current network is calculated by adopting quantitative evaluation indexes, and when the obtained image unmixing accuracy RMSE is greater than or equal to a preset unmixing threshold value, the image unmixing effect of the current network is judged to be poor, and the depth belief network structure in the network training module needs to be adjusted and changed and the network is retrained again; otherwise, calling an output module;
the output module is used for inputting real remote sensing image data to the trained depth belief network in the network training module to obtain a demixing abundance matrix of the real remote sensing image; the real data used in the method is a spectrum library with G wave bands and H ground object types, and G and H are constants larger than 0.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811564932.3ACN109727210B (en) | 2018-12-20 | 2018-12-20 | Based on approximation L0Remote sensing image unmixing method and system of reconstructed deep belief network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811564932.3ACN109727210B (en) | 2018-12-20 | 2018-12-20 | Based on approximation L0Remote sensing image unmixing method and system of reconstructed deep belief network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109727210A CN109727210A (en) | 2019-05-07 |
| CN109727210Btrue CN109727210B (en) | 2021-03-30 |
Family
ID=66296929
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811564932.3AExpired - Fee RelatedCN109727210B (en) | 2018-12-20 | 2018-12-20 | Based on approximation L0Remote sensing image unmixing method and system of reconstructed deep belief network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109727210B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104318245A (en)* | 2014-10-20 | 2015-01-28 | 西安电子科技大学 | Sparse depth network based polarization SAR (Synthetic Aperture Radar) image classification |
| CN104517284A (en)* | 2014-09-29 | 2015-04-15 | 西安电子科技大学 | Polarimetric SAR (synthetic aperture radar) image segmentation based on DBN (deep belief network) |
| CN107145830A (en)* | 2017-04-07 | 2017-09-08 | 西安电子科技大学 | Hyperspectral Image Classification Method Based on Spatial Information Enhancement and Deep Belief Network |
| CN107403200A (en)* | 2017-08-10 | 2017-11-28 | 北京亚鸿世纪科技发展有限公司 | Improve the multiple imperfect picture sorting technique of image segmentation algorithm combination deep learning |
| US9940520B2 (en)* | 2015-05-01 | 2018-04-10 | Applied Research LLC. | Automatic target recognition system with online machine learning capability |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104102929B (en)* | 2014-07-25 | 2017-05-03 | 哈尔滨工业大学 | Hyperspectral remote sensing data classification method based on deep learning |
| EP3345129A4 (en)* | 2015-08-31 | 2019-07-24 | Cape Analytics, Inc. | SYSTEMS AND METHODS FOR ANALYZING REMOTE DETECTION IMAGING |
| CN106529458A (en)* | 2016-10-31 | 2017-03-22 | 重庆大学 | Deep neural network space spectrum classification method for high-spectral image |
| CN107194373A (en)* | 2017-06-16 | 2017-09-22 | 河海大学 | A kind of target in hyperspectral remotely sensed image feature extraction and classifying method |
| CN108388863A (en)* | 2018-02-27 | 2018-08-10 | 南昌工程学院 | A kind of hyperspectral remote sensing image mixed pixel decomposition method |
| CN108509380A (en)* | 2018-03-27 | 2018-09-07 | 北京科技大学 | A kind of high spectrum image solution mixing method |
- 2018
- 2018-12-20CNCN201811564932.3Apatent/CN109727210B/ennot_activeExpired - Fee Related
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104517284A (en)* | 2014-09-29 | 2015-04-15 | 西安电子科技大学 | Polarimetric SAR (synthetic aperture radar) image segmentation based on DBN (deep belief network) |
| CN104318245A (en)* | 2014-10-20 | 2015-01-28 | 西安电子科技大学 | Sparse depth network based polarization SAR (Synthetic Aperture Radar) image classification |
| US9940520B2 (en)* | 2015-05-01 | 2018-04-10 | Applied Research LLC. | Automatic target recognition system with online machine learning capability |
| CN107145830A (en)* | 2017-04-07 | 2017-09-08 | 西安电子科技大学 | Hyperspectral Image Classification Method Based on Spatial Information Enhancement and Deep Belief Network |
| CN107403200A (en)* | 2017-08-10 | 2017-11-28 | 北京亚鸿世纪科技发展有限公司 | Improve the multiple imperfect picture sorting technique of image segmentation algorithm combination deep learning |
Non-Patent Citations (1)
| Title |
|---|
| Application of Deep Belief Network to Land Cover Classification Using Hyperspectral Images;Bulent Ayhan等;《International Symposium on Neural Networks 2017》;20170531;第269-276页* |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109727210A (en) | 2019-05-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112232280B (en) | Hyperspectral image classification method based on self-encoder and 3D depth residual error network | |
| Yang et al. | Deephoyer: Learning sparser neural network with differentiable scale-invariant sparsity measures | |
| CN110210313B (en) | Hyperspectral remote sensing image classification method based on multi-scale PCA-3D-CNN (principal component analysis-three dimensional-CNN) space spectrum combination | |
| CN110334765B (en) | Remote sensing image classification method based on attention mechanism multi-scale deep learning | |
| CN112052755A (en) | Semantic convolutional hyperspectral image classification method based on multi-channel attention mechanism | |
| CN110245709A (en) | Based on deep learning and from the 3D point cloud data semantic dividing method of attention | |
| CN112163628A (en) | Method for improving target real-time identification network structure suitable for embedded equipment | |
| CN110232419A (en) | A kind of method of side slope rock category automatic identification | |
| CN109800795A (en) | A kind of fruit and vegetable recognition method and system | |
| CN113435254A (en) | Sentinel second image-based farmland deep learning extraction method | |
| CN116310510A (en) | Hyperspectral image classification method based on small sample deep learning | |
| CN106600595A (en) | Human body characteristic dimension automatic measuring method based on artificial intelligence algorithm | |
| CN111222545B (en) | Image classification method based on linear programming incremental learning | |
| CN115601751B (en) | A Semantic Segmentation Method of Fundus Image Based on Domain Generalization | |
| CN112949738A (en) | Multi-class unbalanced hyperspectral image classification method based on EECNN algorithm | |
| CN117710744A (en) | Hyperspectral image classification method based on band selection feature fusion multi-attention network | |
| CN112101487A (en) | Compression method and device for fine-grained recognition model | |
| CN118778117B (en) | Compressed excitation ResUnet network seismic data denoising method and system based on spectral normalization | |
| CN106529458A (en) | Deep neural network space spectrum classification method for high-spectral image | |
| CN115511051B (en) | Method for improving performance of convolutional neural network based on weight attention | |
| CN103955711B (en) | A kind of mode identification method in imaging spectral target identification analysis | |
| CN116301914A (en) | Convolutional neural network deployment method based on GAP8 microprocessor | |
| Wang et al. | Multiscale fusion network based on global weighting for hyperspectral feature selection | |
| CN109727210B (en) | Based on approximation L0Remote sensing image unmixing method and system of reconstructed deep belief network | |
| CN114782695A (en) | Non-uniform noise removal method based on stepped multi-scale convolutional neural network |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee | Granted publication date:20210330 Termination date:20211220 | |
| CF01 | Termination of patent right due to non-payment of annual fee |