CN110706154B - Image super-resolution method based on hierarchical residual error neural network - Google Patents
Image super-resolution method based on hierarchical residual error neural network Download PDFInfo
- Publication number
- CN110706154B CN110706154B CN201910802779.1A CN201910802779A CN110706154B CN 110706154 B CN110706154 B CN 110706154B CN 201910802779 A CN201910802779 A CN 201910802779A CN 110706154 B CN110706154 B CN 110706154B
- Authority
- CN
- China
- Prior art keywords
- residual error
- module
- wide
- activation
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 33
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 11
- 238000013507 mapping Methods 0.000 claims abstract description 24
- 238000000605 extraction Methods 0.000 claims abstract description 11
- 230000004927 fusion Effects 0.000 claims abstract description 9
- 238000003062 neural network model Methods 0.000 claims abstract description 9
- 238000005070 sampling Methods 0.000 claims abstract description 3
- 230000004913 activation Effects 0.000 claims description 14
- 230000000694 effects Effects 0.000 claims description 13
- 230000006870 function Effects 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 abstract description 3
- 238000003384 imaging method Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 6
- 238000011439 discrete element method Methods 0.000 description 6
- 238000012549 training Methods 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 2
- 230000000295 complement effect Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 101100365548 Caenorhabditis elegans set-14 gene Proteins 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 230000002085 persistent effect Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000009827 uniform distribution Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4053—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
- G06F17/147—Discrete orthonormal transforms, e.g. discrete cosine transform, discrete sine transform, and variations therefrom, e.g. modified discrete cosine transform, integer transforms approximating the discrete cosine transform
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20048—Transform domain processing
- G06T2207/20052—Discrete cosine transform [DCT]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Mathematical Optimization (AREA)
- Molecular Biology (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Pure & Applied Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- Discrete Mathematics (AREA)
- Databases & Information Systems (AREA)
- Algebra (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses an image super-resolution method based on a hierarchical residual error neural network, which adopts a hierarchical residual error learning strategy and a double-domain enhancement module to construct a hierarchical residual error neural network model which is used for learning a complex mapping relation between a low-resolution image and a high-resolution image, and then reconstructs the high-resolution image from the input low-resolution image by utilizing a trained network model; the hierarchical residual error neural network model is composed of a feature extraction layer, a feature mapping layer and a feature fusion layer, wherein the feature extraction layer is composed of convolution modules, the feature mapping layer is composed of a plurality of double-domain enhancement modules in a cascade connection mode, and the feature fusion layer is composed of an up-sampling module and a convolution module. The method can reconstruct better high-resolution images and has the advantages of small network model parameter quantity and high calculation efficiency.
Description
Technical Field
The invention belongs to the technical field of image processing, machine vision and the like, and particularly relates to an image super-resolution method based on a hierarchical residual error neural network.
Background
The resolution of images acquired by the existing imaging equipment such as digital cameras, video cameras and the like is often low, and subsequent image analysis and understanding are seriously influenced.
The traditional image super-resolution method adopts manual characterization feature modeling, is difficult to describe the real complex super-resolution problem, and limits the performance and application range of the method.
The performance of the existing image super-resolution method based on deep learning is improved by increasing the network depth and sharing parameters, but the network model is too huge and is difficult to be used in an actual scene. And the lightweight super-resolution method has poor performance. These factors limit the application of the existing image super-resolution method in the imaging system.
With the continuous progress of imaging science and technology, people have higher and higher requirements on the resolution of images acquired by imaging equipment. However, due to inherent limitations of imaging devices and interference of the imaging environment, the resolution of the actually acquired images is often low, but high resolution images are often required in reality. For example, in the field of security monitoring, the image resolution is determined by the imaging parameter settings such as the pixel size, the exposure time, the focal length, etc., but is affected by noise, aberration, motion blur, illumination variation, etc. As a common software solution, the image super-resolution directly reconstructs a high-resolution image from a degraded low-resolution image without changing the existing image acquisition process, and is widely applied to the fields of image processing, machine vision and the like, in particular to medical images, remote sensing images and security monitoring.
Over the past 40 years, although a number of image super-resolution methods have emerged, including interpolation-based methods, reconstruction-based methods, and learning-based methods, their effectiveness has been less than ideal. In recent years, the method based on deep learning achieves better effect, and the performance of the method exceeds that of the traditional method, thereby bringing great attention to people. The existing image super-resolution method based on deep learning generally improves the performance by increasing the network depth, but still cannot meet the requirements of many application scenarios in reality, for example, from a convolutional neural network SRCNN of 3 layers to a multi-scale deep super-resolution network MDSR of more than 160 layers. The depth and performance of the network models are remarkably improved, but the training difficulty of the deep network models is increased due to gradient explosion or disappearance. The subsequent high-precision super-resolution method VDSR based on the extremely deep convolutional network and the super-resolution method DRCN based on the deep recursive convolutional network improve the network performance by introducing residual learning, but cannot use different network hierarchical information. Then, an image restoration method MemNet based on a persistent memory network utilizes a memory block to construct a network model, but has the defect that a local convolution layer cannot directly access a subsequent network layer. The dense block can extract the multi-layer features of the deep network, but increases the difficulty of network training. The super-resolution method RDN based on the residual dense network improves the convergence of network training by introducing residual learning, but the network depth is larger. The cascade residual network-based lightweight super-resolution method CARN reduces network depth and computational complexity by sacrificing a little performance. Although these methods achieve good results, they lack efficient network models and the performance still cannot meet the requirements of practical applications.
Disclosure of Invention
Aiming at the defects or shortcomings in the prior art, the invention aims to provide an image super-resolution method based on a hierarchical residual neural network, which is used for learning complex mapping between low-resolution and high-resolution images, using a trained network model for high-resolution image reconstruction and promoting popularization and application of a super-resolution technology in the fields of medical images, remote sensing images, security monitoring and the like.
In order to realize the task, the invention adopts the following technical solution:
an image super-resolution method based on a hierarchical residual error neural network is characterized in that a hierarchical residual error learning strategy and a double-domain enhancement module are adopted to construct a hierarchical residual error neural network model which is used for learning a complex mapping relation between a low-resolution image and a high-resolution image, and then the trained network model is utilized to reconstruct the high-resolution image from the input low-resolution image;
the hierarchical residual error neural network model is composed of a feature extraction layer, a feature mapping layer and a feature fusion layer, wherein the feature extraction layer is composed of convolution modules, the feature mapping layer is composed of a plurality of double-domain enhancement modules in a cascade connection mode, and the feature fusion layer is composed of an up-sampling module and a convolution module.
According to the invention, the double-domain enhancement module consists of 3 wide-activation residual error dense blocks, 2 convolution modules, 1 inverse discrete cosine transform and jump connection; the double-domain enhancement module firstly utilizes two cascaded wide activation residual error dense blocks and a single wide activation residual error dense block to respectively map the space domain characteristics and the frequency domain characteristics, secondly utilizes inverse discrete cosine transform to convert the frequency domain output characteristics into a space domain and to splice the space domain output characteristics, then sends the space domain output characteristics into a next-stage double-domain enhancement module through a convolution module, and finally utilizes outer-layer network residual error learning to improve the network performance and the characteristic mapping effect;
the wide-activation residual error dense block is composed of 3 wide-activation residual error modules, 1 convolution module and jump connection, more features are extracted by the wide-activation residual error dense block through a feedforward mode and the dense block, the convolution module with the size of 1 x 1 is adopted to ensure that the number of input feature channels is consistent with the number of output feature channels, and therefore the residual error learning effect of the middle-layer network is improved;
the wide-activation residual error module is composed of 2 convolution modules, 1 exponential linear unit and jump connection, the wide-activation residual error module strengthens residual error learning effect of an inner network by using a wide-activation mode, and utilizes an exponential linear activation function to correct mapping characteristics to realize characteristic information extraction.
Compared with the prior art, the image super-resolution method based on the hierarchical residual error neural network can reconstruct better high-resolution images, and has the advantages of smaller network model parameter quantity and high calculation efficiency. The double-domain enhancement module can effectively mine complementary information of an airspace and a frequency domain, a layering residual learning strategy improves forward propagation and backward propagation of network signals, and the designed flexible and efficient layering residual neural network model can recover more fine structures and improve the image super-resolution reconstruction effect. In addition, the experimental result verifies the effectiveness and superiority of the hierarchical residual error learning strategy and the double-domain enhancement module.
Drawings
FIG. 1 is a flow diagram of a hierarchical residual neural network model;
FIG. 2 is a network architecture diagram of a dual-domain enhancement module (diagram a), a wide-activation residual dense block (diagram b), and a wide-activation residual module (diagram c);
the present invention will be described in further detail with reference to the following drawings and examples.
Detailed Description
The embodiment provides an image super-resolution method based on a hierarchical residual error neural network, and relates to three parts of a network structure, image preprocessing, implementation details and the like. The method adopts a layered residual error learning strategy and a double-domain enhancement module to construct a layered residual error neural network model which is used for learning a complex mapping relation between a low-resolution image and a high-resolution image, and then a high-resolution image is reconstructed from an input low-resolution image by utilizing the trained network model.
Fig. 1 is a network structure diagram of an image super-resolution reconstruction model, which is composed of a feature extraction layer, a feature mapping layer and a feature fusion layer, wherein the feature extraction layer is composed of a convolution (Conv) module, the feature mapping layer is composed of a plurality of cascaded dual-Domain Enhancement Modules (DEMs), and the feature fusion layer is composed of an upsampling (upsampling) module and a convolution (Conv) module.
1) Network architecture
The image super-resolution reconstruction model adopts five operations: forward Discrete Cosine Transform (DCT), Inverse Discrete Cosine Transform (IDCT), convolution (Conv), concatenation (Concat), and Exponential Linear Unit (ELU) as an activation function. Among them, DCT and IDCT are used for interconversion of spatial and frequency domains. As can be seen from FIG. 1, for an input low-resolution image, the image super-resolution reconstruction model is built on a layerBased on a secondary residual error learning strategy, firstly, the size k is utilized1×k1The convolution module extracts features, then the extracted spatial domain features and frequency domain features are sent into a feature mapping layer formed by a plurality of double-domain enhancement modules, then the mapping features and the residual errors of input images are sent into a feature fusion layer formed by an upsampling and convolution module, and finally high-resolution images are output, wherein the layered residual error learning strategy respectively utilizes a double-Domain Enhancement Module (DEM), a wide-activation residual error dense block (WRDB) and a wide-activation residual error module (WRB) to realize the residual error learning of an outer layer network, a middle layer network and an inner layer network, and the feature mapping effect and the super-resolution network performance are improved.
(1) double-Domain Enhancement Module (DEM)
The dual-Domain Enhancement Module (DEM) is composed of 3 wide-activation residual dense blocks (WRDB), 2 convolution modules (Conv), 1 Inverse Discrete Cosine Transform (IDCT) and hop connections, and its network structure is shown in fig. 2 (a).
The dual-Domain Enhancement Module (DEM) firstly utilizes two cascaded WRDB and a single WRDB to respectively map the spatial domain characteristics and the frequency domain characteristics, secondly utilizes Inverse Discrete Cosine Transform (IDCT) to convert the frequency domain output characteristics into the spatial domain and to splice the spatial domain output characteristics, then sends the spatial domain output characteristics into a next-stage dual-Domain Enhancement Module (DEM) through the convolution module, and finally utilizes outer-layer network residual error learning to improve the network performance and the characteristic mapping effect.
(2) Wide activation residual dense block (WRDB)
The wide activation residual dense block (WRDB) is composed of 3 wide activation residual modules (WRB), 1 convolution module and hop connection, and the network structure is shown in FIG. 2 (b).
The wide-activation residual error dense block (WRDB) extracts more features by utilizing a feed-forward mode and the dense block, and adopts a convolution module with the size of 1 multiplied by 1 to ensure that the number of input feature channels is consistent with the number of output feature channels, so that the residual error learning effect of a middle-layer network is improved.
(3) Wide activation residual block (WRB):
the wide-activation residual block (WRB) is composed of 2 convolution blocks (Conv), 1 Exponential Linear Unit (ELU) and hop-joins, and the network structure thereof is shown in fig. 2 (c).
The wide activation residual error module (WRB) strengthens residual error learning effect of an inner network by using a wide activation mode, and utilizes an exponential linear activation function (ELU) to correct mapping characteristics to realize characteristic information extraction.
2) Image pre-processing
And the training images are preprocessed by adopting a data expansion technology, such as operations of turning, rotating and the like, so that the capacity and diversity of training data samples are increased, and the generalization capability of the model is improved.
3) Implementation details
For the input low resolution image X, the network model described above is intended to be used by learning the mapping function
To reconstruct the super-resolution image Y. Using mean absolute error minimization (i.e., L) between the super-resolution image and the reference image1Norm) to construct a loss function, and a hierarchical residual error learning strategy is proposed to improve the performance and convergence of the network.
To minimize the loss function, the present embodiment trains the network model with an adaptive moment estimation (Adam) optimizer, where the Adam parameters are configured as follows:
β1＝0.9，β2＝0.999，eps＝1e-08。
in the present embodiment, the learning rate starts from lr ═ 0.0001, decays by half every 30 ten thousand times, and the initialization of all the weights and deviation parameters follows uniform distribution θ — (k, k); wherein,
cinis the number of input mapping feature channels.
According to the image super-resolution method based on the hierarchical residual neural network of the embodiment, the applicant carries out test experiments with magnification of 2, 3 and 4 respectively on 5 public data sets (Set5, Set14, BSD100 and Urban100) commonly used today, and the experimental results are shown in the following table.
As can be seen from the table, the image super-resolution method based on the hierarchical residual neural network of the embodiment can reconstruct a better high-resolution image, and has smaller network model parameters and high calculation efficiency. The double-domain enhancement module can effectively mine complementary information of an airspace and a frequency domain, a layering residual learning strategy improves forward propagation and backward propagation of network signals, a designed flexible and efficient network model can recover more fine structures, and the image super-resolution reconstruction effect is improved. The experimental result verifies the effectiveness and superiority of the hierarchical residual error learning strategy and the double-domain enhancement module.
Claims (1)
1. An image super-resolution method based on a hierarchical residual error neural network is characterized in that a hierarchical residual error learning strategy and a double-domain enhancement module are adopted to construct a hierarchical residual error neural network model which is used for learning a complex mapping relation between a low-resolution image and a high-resolution image, and then the trained network model is utilized to reconstruct the high-resolution image from the input low-resolution image;
the hierarchical residual error neural network model consists of a feature extraction layer, a feature mapping layer and a feature fusion layer, wherein the feature extraction layer consists of a convolution module, the feature mapping layer consists of a plurality of double-domain enhancement modules in cascade connection, and the feature fusion layer consists of an up-sampling module and a convolution module;
the double-domain enhancement module consists of 3 wide-activation residual error dense blocks, 2 convolution modules, 1 inverse discrete cosine transform and jump connection; the double-domain enhancement module firstly utilizes two cascaded wide activation residual error dense blocks and a single wide activation residual error dense block to respectively map the space domain characteristics and the frequency domain characteristics, secondly utilizes inverse discrete cosine transform to convert the frequency domain output characteristics into a space domain and to splice the space domain output characteristics, then sends the space domain output characteristics into a next-stage double-domain enhancement module through a convolution module, and finally utilizes outer-layer residual error learning to improve the network performance and the characteristic mapping effect;
the wide-activation residual error dense block is composed of 3 wide-activation residual error modules, 1 convolution module and jump connection, more features are extracted by the wide-activation residual error dense block through a feedforward mode and the dense block, the convolution module with the size of 1 x 1 is adopted to ensure that the number of input feature channels is consistent with the number of output feature channels, and therefore the residual error learning effect of the middle-layer network is improved;
the wide-activation residual error module is composed of 2 convolution modules, 1 exponential linear unit and jump connection, the wide-activation residual error module strengthens residual error learning effect of an inner network by using a wide-activation mode, and utilizes an exponential linear activation function to correct mapping characteristics to realize characteristic information extraction.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910802779.1A CN110706154B (en) | 2019-08-28 | 2019-08-28 | Image super-resolution method based on hierarchical residual error neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910802779.1A CN110706154B (en) | 2019-08-28 | 2019-08-28 | Image super-resolution method based on hierarchical residual error neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110706154A CN110706154A (en) | 2020-01-17 |
| CN110706154B true CN110706154B (en) | 2021-08-31 |
Family
ID=69194117
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910802779.1A Active CN110706154B (en) | 2019-08-28 | 2019-08-28 | Image super-resolution method based on hierarchical residual error neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110706154B (en) |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111445392B (en) * | 2020-03-20 | 2023-09-15 | Oppo广东移动通信有限公司 | Image processing method and device, computer readable storage medium and electronic equipment |
| CN113449756B (en) * | 2020-03-26 | 2022-08-16 | 太原理工大学 | Improved DenseNet-based multi-scale image identification method and device |
| CN112330540A (en) * | 2020-11-08 | 2021-02-05 | 哈尔滨理工大学 | Image super-resolution method based on multi-scale optimization |
| CN112465701B (en) * | 2020-11-26 | 2023-08-29 | 上海健康医学院 | Deep learning super-resolution reconstruction method of microscopic image, medium and electronic equipment |
| CN113902618B (en) * | 2021-10-09 | 2024-03-29 | 普达迪泰（天津）智能装备科技有限公司 | Image super-resolution algorithm based on multi-modal spatial filtering |
| CN114331842B (en) * | 2021-12-28 | 2024-04-05 | 武汉大学 | DEM super-resolution reconstruction method combining topographic features |
| CN114549673B (en) * | 2022-02-25 | 2023-06-23 | 电子科技大学 | Image compression method for preprocessing image based on learning frequency domain information |
| CN117291802B (en) * | 2023-09-27 | 2024-08-02 | 华北电力大学（保定） | Image super-resolution reconstruction method and system based on composite network structure |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3319039A1 (en) * | 2016-11-07 | 2018-05-09 | UMBO CV Inc. | A method and system for providing high resolution image through super-resolution reconstruction |
| CN108921786A (en) * | 2018-06-14 | 2018-11-30 | 天津大学 | Image super-resolution reconstructing method based on residual error convolutional neural networks |
| CN109636769A (en) * | 2018-12-18 | 2019-04-16 | 武汉大学 | EO-1 hyperion and Multispectral Image Fusion Methods based on the intensive residual error network of two-way |
| CN109829855A (en) * | 2019-01-23 | 2019-05-31 | 南京航空航天大学 | A kind of super resolution ratio reconstruction method based on fusion multi-level features figure |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104778671B (en) * | 2015-04-21 | 2017-09-22 | 重庆大学 | A kind of image super-resolution method based on SAE and rarefaction representation |
-
2019
- 2019-08-28 CN CN201910802779.1A patent/CN110706154B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3319039A1 (en) * | 2016-11-07 | 2018-05-09 | UMBO CV Inc. | A method and system for providing high resolution image through super-resolution reconstruction |
| CN108921786A (en) * | 2018-06-14 | 2018-11-30 | 天津大学 | Image super-resolution reconstructing method based on residual error convolutional neural networks |
| CN109636769A (en) * | 2018-12-18 | 2019-04-16 | 武汉大学 | EO-1 hyperion and Multispectral Image Fusion Methods based on the intensive residual error network of two-way |
| CN109829855A (en) * | 2019-01-23 | 2019-05-31 | 南京航空航天大学 | A kind of super resolution ratio reconstruction method based on fusion multi-level features figure |
Non-Patent Citations (2)
| Title |
|---|
| 基于残差学习的图像超分辨率重建;魏雪琦;《中国优秀硕士学位论文全文数据库 信息科技辑》;20190215(第02期);I138-2034 * |
| 基于深度学习的图像超分辨率重建算法研究;段君毅;《中国优秀硕士学位论文全文数据库 信息科技辑》;20190715(第07期);I138-1259 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110706154A (en) | 2020-01-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110706154B (en) | Image super-resolution method based on hierarchical residual error neural network | |
| CN115222601A (en) | Image super-resolution reconstruction model and method based on residual mixed attention network | |
| Saragadam et al. | Miner: Multiscale implicit neural representation | |
| CN111080541B (en) | Color image denoising method based on bit layering and attention fusion mechanism | |
| CN107133919A (en) | Time dimension video super-resolution method based on deep learning | |
| CN109949222B (en) | Image super-resolution reconstruction method based on semantic graph | |
| CN109949217B (en) | Video super-resolution reconstruction method based on residual learning and implicit motion compensation | |
| CN112529776B (en) | Training method of image processing model, image processing method and device | |
| CN113379606B (en) | Face super-resolution method based on pre-training generation model | |
| Liu et al. | Single image super resolution techniques based on deep learning: Status, applications and future directions | |
| CN111462208A (en) | Non-supervision depth prediction method based on binocular parallax and epipolar line constraint | |
| WO2022141660A1 (en) | Neural network-based high-resolution image restoration method and system | |
| Yang et al. | A survey of super-resolution based on deep learning | |
| Luo et al. | Image demosaicing based on generative adversarial network | |
| CN114549361B (en) | Image motion blur removing method based on improved U-Net model | |
| Ai et al. | Single image super-resolution via residual neuron attention networks | |
| CN109272450B (en) | Image super-resolution method based on convolutional neural network | |
| Shuai et al. | Accurate image super-resolution using cascaded multi-column convolutional neural networks | |
| CN117474764A (en) | High-resolution reconstruction method for remote sensing image under complex degradation model | |
| Sun et al. | Two-stage deep single-image super-resolution with multiple blur kernels for Internet of Things | |
| CN117593199A (en) | Double-flow remote sensing image fusion method based on Gaussian prior distribution self-attention | |
| CN116309066A (en) | Super-resolution imaging method and device | |
| Shin et al. | LoGSRN: Deep super resolution network for digital elevation model | |
| CN116128722A (en) | Image super-resolution reconstruction method and system based on frequency domain-texture feature fusion | |
| CN106846286B (en) | Video super-resolution algorithm for reconstructing based on a variety of complementary priori |
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 |