CN109615577A - High spectrum image super-resolution processing method based on convolutional network - Google Patents
High spectrum image super-resolution processing method based on convolutional network Download PDFInfo
- Publication number
- CN109615577A CN109615577A CN201811261452.XA CN201811261452A CN109615577A CN 109615577 A CN109615577 A CN 109615577A CN 201811261452 A CN201811261452 A CN 201811261452A CN 109615577 A CN109615577 A CN 109615577A
- Authority
- CN
- China
- Prior art keywords
- resolution
- network
- image
- indicate
- spectrum 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4053—Super resolution, i.e. output image resolution higher than sensor resolution
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of high spectrum image super-resolution processing method based on convolutional network, the high-resolution high spectrum image resolution ratio for mainly solving the problems, such as that the prior art generates is lower, and implementation is: first by collected low resolution high spectrum image and high-resolution multispectral image composing training sample and test sample；Then the convolutional network being made of reasoning sub-network and generation sub-network is constructed；Training convolutional network is removed with training sample, and by maximizing the low joint likelihood function for differentiating high spectrum image and high-resolution multispectral image, obtains APPROXIMATE DISTRIBUTION and the true distribution highest convolutional network of similarity；Finally test sample is input in the convolutional network after training, processing is optimized to the high-resolution high spectrum image of generation, obtains final high-resolution high spectrum image.The present invention improves the resolution ratio for generating high-resolution high spectrum image, can be used for medical diagnosis, remote sensing, computer vision and monitoring by the convolutional neural networks using deep layer.
Description
Technical field
The invention belongs to technical field of image processing, further relate to Image Super-resolution processing method, can be used for medical treatment
Diagnosis, remote sensing, computer vision and monitoring.
Background technique
Super-resolution optical image can overcome the limitation of low resolution optical image, and in such as medical diagnosis, remote sensing, meter
Good result is shown in many applications such as calculation machine vision and monitoring.High-resolution optical imagery is obtained, most directly
Method is using high-resolution image sensors, but due to the limitation of sensor and optical device manufacturing process and cost, very
It is difficult to realize in more occasions and large scale deployment.Therefore, using existing equipment, High-Resolution Map is obtained by super-resolution technique
As having important practical significance.In order to break the limitation of imaging system intrinsic resolution, optical imagery is improved from algorithm angle
The technology of resolution ratio is just come into being.Existing Image Super-resolution mainly has these types: the side based on traditional interpolation theory
Method, based on model/reconstruction method and based on the method for machine learning.
Paper " the Coupled deep that K.Zeng, J.Yu, R.Wang, C.Li, D.Tao et al. are delivered at it
autoencoder for single image super-resolution.”(IEEE Transactions on
Cybernetics, 2015, pages 1-11.) a kind of single width optical imagery super-resolution based on deep learning network is disclosed in
Processing method.This method learns high-resolution/low resolution image block pair intrinsic representation from encoding model using coupling depth, together
When obtain the mapping of low resolution image block to full resolution pricture block, ultimately produce full resolution pricture block, and generate former low resolution figure
As corresponding full resolution pricture.This method shows relatively good performance in reconstruct, and have in feedforward network test compared with
Fast speed.But, however it remains shortcoming be: this method do not account for hidden variable in model it is implicit it is uncertain because
Element, therefore affect the resolution ratio of the full resolution pricture of generation.
(patent application is special in a kind of patent document " image super-resolution method " of its application for Shenzhen Graduate School of Tsinghua University
Benefit number: 201510338958.6, publication number: CN10499240A) in propose a kind of image indicated based on cluster and collaboration super
Resolving method.This method is clustered when organizing cluster centre neighbour, and to search space, and the local geometric that data are utilized is special
Property, to more accurately restore the high-frequency information of high-definition picture, obtain higher-quality high-definition picture, this method energy
It is enough to determine cluster centre, and the statistical property of available each cluster centre in cluster, further to add
Fast super-resolution speed provides may.But, however it remains shortcoming be: this method is due to being the probability using a shallow-layer
Model, and the information on low resolution image surface is only utilized, do not account for the implicit letter of high-resolution/low resolution image block centering
Breath, thus the resolution ratio that more information carry out sophisticated image can not be generated, affect the resolution of the full resolution pricture ultimately produced
Rate.
Summary of the invention
It is an object of the invention to overcome the defect of the above-mentioned prior art, a kind of EO-1 hyperion based on convolutional network is proposed
Image Super-resolution processing method improves the resolution ratio of finally obtained full resolution pricture using the convolutional network structure of deep layer.
To achieve the above object, include the following:
(1) low resolution high spectrum image L and high-resolution multispectral image H are obtained from Harvard image set；
It (2), will be with training sample not using low resolution high spectrum image L and high-resolution multispectral image H as training sample
Same low resolution high spectrum image L and high-resolution multispectral image H is as test sample；
(3) convolutional network is constructed；
Reasoning sub-network is arranged in (3a), which is to be by inputtingHidden variable isTwo convolutional neural networks
The variation autoencoder network of CNN composition, whereinIndicate the low ith pixel point for differentiating high spectrum image L,Indicate low point
Distinguish i-th of hidden variable of high spectrum image L；
(3b) setting generates sub-network, which is to be by inputtingHidden variable isTwo convolutional neural networks
The variation autoencoder network of CNN composition, whereinIndicate j-th of pixel of high-resolution multispectral image H,Indicate high score
Distinguish j-th of hidden variable of multispectral image H；
(4) training convolutional network:
(4a) stochastical sampling from the Gaussian Profile that mean value is 0, variance is 0.01, using the array of stochastical sampling as convolution
The initiation parameter of network；
All low resolution high spectrum image L in training sample are inputted the reasoning sub-networks in convolutional networks by (4b), together
When all high-resolution multispectral image H in training sample are input to the generation sub-network in convolutional network；
(4c) is obtained by maximizing the low joint likelihood function for differentiating high spectrum image L and high-resolution multispectral image H
APPROXIMATE DISTRIBUTION and the true distribution highest reasoning sub-network of similarity and generation sub-network；
(4d) utilizes batch stochastic gradient descent method, to (4c) obtained reasoning sub-network and generates the institute in sub-network
There is parameter to be iterated update, obtain updated reasoning sub-network and generates sub-network；
(5) image optimization is handled:
All low resolution high spectrum image L in test sample are input in updated reasoning sub-network by (5a), together
When all high-resolution multispectral image H in test sample are input in updated generation sub-network, it is high to generate high-resolution
Spectrum picture block, and arranged；
(5b) averages to the pixel value of the lap between image block after arrangement, the high-resolution bloom reconstructed
Spectrogram picture；
(5c) optimizes processing to the high-resolution high spectrum image of reconstruct, obtains final high-resolution high spectrum image.
The present invention has the advantage that compared with prior art
The present invention due to constructing by reasoning sub-network and generating the convolutional network with deep structure that forms of sub-network,
More image informations can be extracted, are overcome in the prior art due to having used the network of shallow-layer to carry out oversubscription to image pattern
It distinguishes processing, leads to not that corresponding high-resolution bloom can only be generated with partial information using all information for including in original image
The deficiency of spectrogram picture effectively raises high score so that the information that the high-resolution high spectrum image that the present invention generates includes is more
Distinguish the resolution ratio of high spectrum image.
Detailed description of the invention
Fig. 1 is implementation flow chart of the invention；
Fig. 2 is the low resolution high spectrum image inputted in the present invention；
Fig. 3 is the high-resolution multispectral image inputted in the present invention；
Fig. 4 is the final high-resolution high spectrum image obtained using the present invention.
Specific embodiment
The present invention will be further described with reference to the accompanying drawing.
Referring to Fig.1, specific implementation step of the invention is as follows.
Step 1. Image Acquisition generates sample.
(1a) obtains low resolution high spectrum image L and high-resolution multispectral image H from Harvard image set；
(1b), will be with training sample not using low resolution high spectrum image L and high-resolution multispectral image H as training sample
Same low resolution high spectrum image L and high-resolution multispectral image H is as test sample.
Step 2. constructs convolutional network.
Reasoning sub-network is arranged in (2a), which is to be by inputtingHidden variable isTwo convolutional neural networks
The variation autoencoder network of CNN composition, whereinIndicate the low ith pixel point for differentiating high spectrum image L,Indicate low point
Distinguish i-th of hidden variable of high spectrum image L；
The likelihood function of reasoning sub-network is as follows:
Wherein, pθ(|) indicates the conditional probability distribution of reasoning sub-network,It indicates low and differentiates high spectrum image L's
Ith pixel point,Indicate low i-th of hidden variable for differentiating high spectrum image L,Indicate Gaussian Profile, μpWithTable
Show withRelated nonlinear function, I indicate unit matrix；
(2b) setting generates sub-network, which is to be by inputtingHidden variable isTwo convolutional neural networks
The variation autoencoder network of CNN composition, whereinIndicate j-th of pixel of high-resolution multispectral image H,Indicate high score
Distinguish j-th of hidden variable of multispectral image H；
The likelihood function for generating sub-network is as follows:
Wherein, pθ(|) indicates to generate the conditional probability distribution of sub-network,Indicate high-resolution multispectral image H's
J-th of pixel,Indicate j-th of hidden variable of high-resolution multispectral image H, μpWithIndicate withIt is related non-linear
Function.
Step 3. training convolutional network.
(3a) stochastical sampling from the Gaussian Profile that mean value is 0, variance is 0.01, using the array of stochastical sampling as convolution
The initiation parameter of network；
All low resolution high spectrum image L in training sample are inputted the reasoning sub-networks in convolutional networks by (3b), together
When all high-resolution multispectral image H in training sample are input to the generation sub-network in convolutional network；
(3c) is obtained by maximizing the low joint likelihood function for differentiating high spectrum image L and high-resolution multispectral image H
APPROXIMATE DISTRIBUTION and the true distribution highest reasoning sub-network of similarity and generation sub-network；
The joint likelihood function formula is as follows:
Wherein, log () indicates denary logarithm operation, and p () indicates joint probability distribution,Table
Show that image dimension is tieed up for a × b, the low resolution high spectrum image that spectrum dimension is C,Indicate image dimension for A × B dimension, spectrum
Dimension is the high-resolution multispectral image of c,Indicate that variation lower limit function, θ indicate the parameter of convolutional network, α indicates to generate son
The parameter of network, φlIndicate the low parameter for differentiating high spectrum image, φhIndicate the parameter of high-resolution multispectral image,Table
Show summation symbol,Expression APPROXIMATE DISTRIBUTION is at a distance from being really distributed, qφ() indicates APPROXIMATE DISTRIBUTION；
(3d) utilizes batch stochastic gradient descent method, to (4c) obtained reasoning sub-network and generates the institute in sub-network
There is parameter to be iterated update, obtain updated reasoning sub-network and generates sub-network；
Step 4. image optimization.
All low resolution high spectrum image L in test sample are input in updated reasoning sub-network by (4a), together
When all high-resolution multispectral image H in test sample are input in updated generation sub-network, it is high to generate high-resolution
Spectrum picture block, and arranged；
(4b) averages to the pixel value of the lap between image block after arrangement, the high-resolution bloom reconstructed
Spectrogram picture；
(4c) optimizes processing to the high-resolution high spectrum image of reconstruct, obtains final high-resolution high spectrum image:
In formula,Indicate the pixel value of final high-resolution high spectrum image,Expression is minimized function
When, the pixel value Y of high-resolution multispectral image in test seth, | | | | indicate norm operation, S indicates down-sampling operation, H table
Show the linear smoothing filtering operation that filter weights are selected according to the shape of Gaussian function, YlIndicate the low resolution bloom in test set
The pixel value of spectrogram picture, YhIndicate that the high-resolution high spectrum image of reconstruct, c indicate constant 0.1,Indicate the quadratic power of two norms
Operation.
Effect of the invention can be further illustrated by following emulation.
1. simulated conditions:
L-G simulation test of the invention is Intel (R) Core (TM) i5-6500CPU, the memory 8GB in dominant frequency 3.2GHz
It is carried out under hardware environment and software environment based on Python2.5.
2. emulation content:
Emulation 1, with the present invention to low resolution high spectrum image shown in Fig. 2 from Harvard image set and such as Fig. 3 institute
The high-resolution multispectral image shown carries out Image Super-resolution emulation, obtains final high-resolution high spectrum image, as shown in Figure 4.
The present invention is extracted the spatial information of the low spectral information for differentiating high spectrum image and high-resolution multispectral image as seen from Figure 4,
With preferable resolution ratio.
Emulation 2, with method of the invention, the existing image super-resolution method based on generalized synchronization orthogonal matching pursuit
GSOMP and super-resolution experiment is carried out to Harvard image set based on the sparse image super-resolution method BSR of Bayes, and by root
It is super to compare three kinds of methods progress images as standard of comparison by mean square error RMSE, global error ERGAS and spectral modeling matching SAM
The precision of resolution, when the value of RMSE, ERGAS and SAM are lower, the resolution ratio of image is higher, as a result such as table 1:
RMSE, ERGAS and SAM that table 1.GSOMP, BSR, the method for the present invention obtain in simulations
From 1 result of table: compared with existing GSOMP, BSR image super-resolution method, the root that the present invention obtains is square
Error RMSE, global error ERGAS and spectral modeling matching SAM are smaller, indicate proposed by the present invention this based on convolutional network
High spectrum image super-resolution processing method improves the deficiency of above two method, can improve the resolution ratio of high spectrum image.
Claims (5)
1. a kind of high spectrum image super-resolution processing method based on convolutional network, includes the following:
(1) low resolution high spectrum image L and high-resolution multispectral image H are obtained from Harvard image set；
It (2), will be different from training sample using low resolution high spectrum image L and high-resolution multispectral image H as training sample
Low resolution high spectrum image L and high-resolution multispectral image H are as test sample；
(3) convolutional network is constructed；
Reasoning sub-network is arranged in (3a), which is to be by inputtingHidden variable isTwo convolutional neural networks CNN groups
At variation autoencoder network, whereinIndicate the low ith pixel point for differentiating high spectrum image L,Indicate low resolution bloom
I-th hidden variable of the spectrogram as L；
(3b) setting generates sub-network, which is to be by inputtingHidden variable isTwo convolutional neural networks CNN groups
At variation autoencoder network, whereinIndicate j-th of pixel of high-resolution multispectral image H,Indicate that high-resolution is more
J-th of hidden variable of spectrum picture H；
(4) training convolutional network:
(4a) stochastical sampling from the Gaussian Profile that mean value is 0, variance is 0.01, using the array of stochastical sampling as convolutional network
Initiation parameter；
All low reasoning sub-networks differentiated in high spectrum image L input convolutional network in training sample simultaneously will by (4b)
All high-resolution multispectral image H in training sample are input to the generation sub-network in convolutional network；
(4c) obtains approximation by maximizing the low joint likelihood function for differentiating high spectrum image L and high-resolution multispectral image H
It is distributed and is really distributed the highest reasoning sub-network of similarity and generation sub-network；
(4d) utilizes batch stochastic gradient descent method, to (4c) obtained reasoning sub-network and generates all ginsengs in sub-network
Number is iterated update, obtains updated reasoning sub-network and generates sub-network；
(5) image optimization is handled:
All low resolution high spectrum image L in test sample are input in updated reasoning sub-network by (5a), simultaneously will
All high-resolution multispectral image H in test sample are input in updated generation sub-network, generate high-resolution EO-1 hyperion
Image block, and arranged；
(5b) averages to the pixel value of the lap between image block after arrangement, the high-resolution high-spectrum reconstructed
Picture；
(5c) optimizes processing to the high-resolution high spectrum image of reconstruct, obtains final high-resolution high spectrum image.
2. the method according to claim 1, wherein the likelihood function of reasoning sub-network is as follows in (3a):
Wherein, pθ(|) indicates the conditional probability distribution of reasoning sub-network,Indicate low i-th for differentiating high spectrum image L
Pixel,Indicate low i-th of hidden variable for differentiating high spectrum image L,Indicate Gaussian Profile, μpWithIndicate withRelated nonlinear function, I indicate unit matrix.
3. the method according to claim 1, wherein the likelihood function for generating sub-network in (3b) is as follows:
Wherein, pθ(|) indicates to generate the conditional probability distribution of sub-network,Indicate j-th of high-resolution multispectral image H
Pixel,Indicate j-th of hidden variable of high-resolution multispectral image H, μpWithIndicate withRelated nonlinear function.
4. the method according to claim 1, wherein joint likelihood function formula described in (4c) is as follows:
Wherein, log () indicates denary logarithm operation, and p () indicates joint probability distribution,Indicate image
Dimension is that a × b is tieed up, the low resolution high spectrum image that spectrum dimension is C,Image dimension is indicated as A × B dimension, spectrum dimension is c
High-resolution multispectral image,Indicate that variation lower limit function, θ indicate the parameter of convolutional network, α indicates to generate sub-network
Parameter, φlIndicate the low parameter for differentiating high spectrum image, φhIndicate the parameter of high-resolution multispectral image,Indicate summation
Symbol,Expression APPROXIMATE DISTRIBUTION is at a distance from being really distributed, qφ() indicates APPROXIMATE DISTRIBUTION.
5. the method according to claim 1, wherein carrying out in (5c) to reconstruct high-resolution high spectrum image is excellent
Change processing, refers to the operation carried out using following formula:
Wherein,Indicate the pixel value of final high-resolution high spectrum image,When expression is minimized function,
The pixel value Y of high-resolution multispectral image in test seth, | | | | indicate norm operation, S indicates down-sampling operation, and H indicates root
According to the linear smoothing filtering operation of the shape selection filter weights of Gaussian function, YlIndicate the low resolution high-spectrum in test set
The pixel value of picture, YhIndicate that the high-resolution high spectrum image of reconstruct, c indicate constant 0.1,Indicate the quadratic power behaviour of two norms
Make.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811261452.XA CN109615577B (en) | 2018-10-26 | 2018-10-26 | Hyperspectral image super-resolution processing method based on convolution network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811261452.XA CN109615577B (en) | 2018-10-26 | 2018-10-26 | Hyperspectral image super-resolution processing method based on convolution network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109615577A true CN109615577A (en) | 2019-04-12 |
| CN109615577B CN109615577B (en) | 2023-01-06 |
Family
ID=66002420
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811261452.XA Active CN109615577B (en) | 2018-10-26 | 2018-10-26 | Hyperspectral image super-resolution processing method based on convolution network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109615577B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110109110A (en) * | 2019-04-26 | 2019-08-09 | 西安电子科技大学 | Based on the optimal variation of priori from the HRRP target identification method of code machine |
| CN112184560A (en) * | 2020-12-02 | 2021-01-05 | 南京理工大学 | Hyperspectral image super-resolution optimization method based on deep closed-loop neural network |
| CN112288627A (en) * | 2020-10-23 | 2021-01-29 | 武汉大学 | Recognition-oriented low-resolution face image super-resolution method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017219263A1 (en) * | 2016-06-22 | 2017-12-28 | 中国科学院自动化研究所 | Image super-resolution enhancement method based on bidirectional recursion convolution neural network |
| CN107622476A (en) * | 2017-08-25 | 2018-01-23 | 西安电子科技大学 | Image Super-resolution processing method based on generative probabilistic model |
| CN108376386A (en) * | 2018-03-23 | 2018-08-07 | 深圳天琴医疗科技有限公司 | A kind of construction method and device of the super-resolution model of image |
-
2018
- 2018-10-26 CN CN201811261452.XA patent/CN109615577B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017219263A1 (en) * | 2016-06-22 | 2017-12-28 | 中国科学院自动化研究所 | Image super-resolution enhancement method based on bidirectional recursion convolution neural network |
| CN107622476A (en) * | 2017-08-25 | 2018-01-23 | 西安电子科技大学 | Image Super-resolution processing method based on generative probabilistic model |
| CN108376386A (en) * | 2018-03-23 | 2018-08-07 | 深圳天琴医疗科技有限公司 | A kind of construction method and device of the super-resolution model of image |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110109110A (en) * | 2019-04-26 | 2019-08-09 | 西安电子科技大学 | Based on the optimal variation of priori from the HRRP target identification method of code machine |
| CN110109110B (en) * | 2019-04-26 | 2023-06-16 | 西安电子科技大学 | HRRP target identification method based on priori optimal variation self-encoder |
| CN112288627A (en) * | 2020-10-23 | 2021-01-29 | 武汉大学 | Recognition-oriented low-resolution face image super-resolution method |
| CN112288627B (en) * | 2020-10-23 | 2022-07-05 | 武汉大学 | Recognition-oriented low-resolution face image super-resolution method |
| CN112184560A (en) * | 2020-12-02 | 2021-01-05 | 南京理工大学 | Hyperspectral image super-resolution optimization method based on deep closed-loop neural network |
| CN112184560B (en) * | 2020-12-02 | 2021-03-26 | 南京理工大学 | Hyperspectral image super-resolution optimization method based on deep closed-loop neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109615577B (en) | 2023-01-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110232341B (en) | Semi-supervised learning image identification method based on convolution-stacking noise reduction coding network | |
| CN109146831A (en) | Remote sensing image fusion method and system based on double branch deep learning networks | |
| CN114092832B (en) | High-resolution remote sensing image classification method based on parallel hybrid convolutional network | |
| CN106022355B (en) | High spectrum image sky based on 3DCNN composes joint classification method | |
| CN108921926A (en) | A kind of end-to-end three-dimensional facial reconstruction method based on single image | |
| CN108304357A (en) | A kind of Chinese word library automatic generation method based on font manifold | |
| CN109325513B (en) | Image classification network training method based on massive single-class images | |
| CN106780546B (en) | The personal identification method of motion blur encoded point based on convolutional neural networks | |
| CN109615577A (en) | High spectrum image super-resolution processing method based on convolutional network | |
| CN111401156B (en) | Image identification method based on Gabor convolution neural network | |
| CN112102165B (en) | Light field image angular domain super-resolution system and method based on zero sample learning | |
| CN113066037B (en) | Multispectral and full-color image fusion method and system based on graph attention machine system | |
| CN110992366A (en) | Image semantic segmentation method and device and storage medium | |
| WO2020168648A1 (en) | Image segmentation method and device, and computer-readable storage medium | |
| CN112634149A (en) | Point cloud denoising method based on graph convolution network | |
| CN107622476B (en) | Image Super-resolution processing method based on generative probabilistic model | |
| CN109584194B (en) | Hyperspectral image fusion method based on convolution variation probability model | |
| CN116933141B (en) | Multispectral laser radar point cloud classification method based on multicore graph learning | |
| CN112686830A (en) | Super-resolution method of single depth map based on image decomposition | |
| CN117036901A (en) | Small sample fine adjustment method based on visual self-attention model | |
| CN117094925A (en) | Pig body point cloud completion method based on point agent enhancement and layer-by-layer up-sampling | |
| CN113628111B (en) | Hyperspectral image super-resolution method based on gradient information constraint | |
| CN115908600A (en) | Massive image reconstruction method based on prior regularization | |
| CN106530389B (en) | Stereo reconstruction method based on medium-wave infrared facial image | |
| Li et al. | Research on visual‐tactile cross‐modality based on generative adversarial 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 |