CN110175953A - A kind of image super-resolution method and system - Google Patents
A kind of image super-resolution method and system Download PDFInfo
- Publication number
- CN110175953A CN110175953A CN201910439532.8A CN201910439532A CN110175953A CN 110175953 A CN110175953 A CN 110175953A CN 201910439532 A CN201910439532 A CN 201910439532A CN 110175953 A CN110175953 A CN 110175953A
- Authority
- CN
- China
- Prior art keywords
- image
- feature
- module
- neural networks
- convolutional neural
- 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
- 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
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Abstract
It includes: using image to be processed as the input of convolutional neural networks oversubscription model that the present invention, which provides a kind of image super-resolution method and system, the program, and convolutional neural networks oversubscription model is made of four sequentially connected execution modules；First execution module handles image to be processed, obtains the first processing image；Second execution module handles the first processing image, and output includes second processing image；Third execution module handles second processing image, and output third handles image；4th execution module handles third processing image, exports super-resolution image.Based on the present invention, convolutional neural networks oversubscription model is that weighted feature is arranged in image to be processed, pass through the study to weighted feature, determine the important feature in image to be processed, and superresolution processing is carried out according to important feature, to improve the feature representation ability of convolutional neural networks oversubscription model, so that the details quality of obtained super-resolution image greatly improves after superresolution processing.
Description
Technical field
The present invention relates to field of artificial intelligence more particularly to a kind of image super-resolution method and systems.
Background technique
Image Super-resolution be in the computer vision systems such as video monitoring, logistics, recognition of face one it is very important pre-
Processing links.Image Super-resolution can reconstruct super-resolution image from low-resolution image, be wanted with meeting the browsing of user
It asks.Have benefited from the powerful non-linear expression's ability of convolutional neural networks (Convolutional Neural Networks, CNN),
Convolutional neural networks oversubscription model realization be can use to the superresolution processing of low-resolution image, obtain super-resolution image.
Currently, the superresolution processing performance of existing convolutional neural networks oversubscription model, often only and convolutional neural networks
The depth or width of structure are related.However, passing through super-resolution brought by the depth for increasing convolutional neural networks structure or width
Process performance promotion effect is extremely limited, is unable to satisfy user to the high standards of super-resolution image.So existing volume
The super-resolution image that product neural network oversubscription model generates still has more shortcoming.
Summary of the invention
In view of this, the embodiment of the present invention provides a kind of image super-resolution method and system, to solve via existing volume
The unconspicuous problem of image detail effect for the super-resolution image that product neural network oversubscription model generates.
To achieve the above object, the embodiment of the present invention provides the following technical solutions:
First aspect of the embodiment of the present invention discloses a kind of image super-resolution method, and described image ultra-resolution method includes:
Using image to be processed as the input of the convolutional neural networks oversubscription model constructed in advance, the convolutional neural networks
Oversubscription model is made of four sequentially connected execution modules, and the second execution module is by being superimposed and being embedded in second order channel attention mould
The residual error module composition of block；
The image to be processed is handled via the first execution module in the convolutional neural networks oversubscription model,
Obtain input of the first processing image as second execution module, the size of the first processing image with it is described to be processed
The size of image is identical；
Feature extraction is carried out to the first processing image using second execution module and characteristic processing, output include
Input of the second processing image of weighted feature as the third execution module in the convolutional neural networks oversubscription model；
Size preset ratio system based on the third execution module and pre-set input picture and output image
Number, handles the second processing image, the third processing image that size meets preset ratio coefficient is obtained, by described the
Input of the three processing images as the 4th execution module in the convolutional neural networks oversubscription model；
Mapping processing, the corresponding image to be processed of output are carried out to third processing image via the 4th execution module
Super-resolution image.
Optionally, in above-mentioned image super-resolution method, the mistake of the convolutional neural networks oversubscription model constructed in advance
Journey, comprising:
Training set is constructed, the training set includes low-resolution image and height corresponding with the low-resolution image
Image in different resolution；
The low-resolution image is input in preset convolutional neural networks model and carries out feature extraction, feature amplification
And Feature Mapping, the image that obtains that treated；
After the low-resolution image, high-definition picture corresponding with the low-resolution image and processing
Image, using preset loss function and the optimization algorithm training preset convolutional neural networks model, until described pre-
If convolutional neural networks model export corresponding with low-resolution image high-definition picture, determine currently to train and obtain
Convolutional neural networks model be convolutional neural networks oversubscription model；
Wherein, the convolutional neural networks oversubscription model is made of four sequentially connected execution modules, and second executes mould
Block pays attention to the residual error module composition of power module by being superimposed and being embedded in second order channel.
Optionally, in above-mentioned image super-resolution method, the residual error for being superimposed and being embedded in second order channel and pay attention to power module
The process of the second execution module of module composition, comprising:
By preset second order channel attention Module-embedding into residual error module, the residual error mould with weighted feature is obtained
Block；
The number of residual error module with weighted feature needed for determining the second execution module of building；
Each residual error module with weighted feature is stacked gradually, second execution module is obtained.
Optionally, described by second order channel attention Module-embedding to residual error module in above-mentioned image super-resolution method
In, obtain the residual error module with weighted feature, comprising:
The first convolutional layer, active coating, the second convolutional layer, the first residual unit, second order channel note are successively connected in sequence
Power module of anticipating and the second residual unit obtain the residual error module with weighted feature.
Optionally, in above-mentioned image super-resolution method, the default second order channel pays attention to the process of power module, comprising:
Based on the first processing image as input, the fisrt feature of any layer in convolutional neural networks convolutional layer is obtained；
Mapping processing is carried out to the fisrt feature according to matrix recombination method, obtains second feature；
Transposition based on the fisrt feature and second feature, is calculated sample variance matrix；
The sample variance matrix is normalized, covariance matrix is obtained；
Based on the covariance matrix, the row mean vector based on depth dimension is calculated；
Dimensionality reduction study is successively carried out to the row mean vector based on depth dimension and rises dimension study, obtains the first power
Weight；
First weight is normalized, the second weight is obtained；
Based on the fisrt feature and the second weight, weighted feature is obtained；
Power module is paid attention to using weighted feature building second order channel.
Second aspect of the embodiment of the present invention discloses a kind of image super-resolution system, and described image super-resolution system includes:
Input unit, for using image to be processed as the input of the convolutional neural networks oversubscription model constructed in advance, institute
It states convolutional neural networks oversubscription model to be made of four sequentially connected execution modules, the second execution module is by being superimposed and being embedded in two
Rank channel pays attention to the residual error module composition of power module；
First execution unit, for via the first execution module in the convolutional neural networks oversubscription model to it is described to
Processing image is handled, and input of the first processing image as second execution module, the first processing image are obtained
Size it is identical as the size of the image to be processed；
Second execution unit, for using second execution module to it is described first processing image carry out feature extraction with
Characteristic processing, output are held comprising the second processing image of weighted feature as the third in the convolutional neural networks oversubscription model
The input of row module；
Third execution unit, for based on the third execution module and pre-set input picture and output figure
The size preset ratio coefficient of picture, handles the second processing image, obtains size meets preset ratio coefficient
Three processing images, using third processing image as the defeated of the 4th execution module in the convolutional neural networks oversubscription model
Enter；
4th execution unit is exported for carrying out mapping processing to third processing image via the 4th execution module
The super-resolution image of the corresponding image to be processed.
Optionally, in above-mentioned image super-resolution system, the input unit includes:
Construct subelement, for constructing training set, the training set include low-resolution image and with the low resolution
The corresponding high-definition picture of rate image；
Subelement is handled, carries out spy for the low-resolution image to be input in preset convolutional neural networks model
Sign extracts, feature amplification and Feature Mapping, the image that obtains that treated；
Training subelement, for being based on the low-resolution image, high-resolution corresponding with the low-resolution image
Image and treated image utilize preset loss function and the optimization algorithm training preset convolutional neural networks
Model, until the preset convolutional neural networks model exports high-definition picture corresponding with the low-resolution image,
Determine that the convolutional neural networks model that current training obtains is convolutional neural networks oversubscription model；Wherein, the convolutional Neural net
Network oversubscription model is made of four sequentially connected execution modules, and the second execution module is by being superimposed and being embedded in second order channel attention
The residual error module composition of module.
Optionally, in above-mentioned image super-resolution system, further includes:
First construction unit, for preset second order channel attention Module-embedding into residual error module, to be had
The residual error module of weighted feature, the number of the residual error module with weighted feature needed for determining the second execution module of building, according to
It is secondary to stack each residual error module with weighted feature, obtain second execution module.
Optionally, described by second order channel attention Module-embedding to residual error module in above-mentioned image super-resolution system
In, the first construction unit for obtaining the residual error module with weighted feature is specifically used for successively connecting the first convolution in sequence
Layer, active coating, the second convolutional layer, second order channel pay attention to power module and residual error, obtain the residual error module with weighted feature.
Optionally, in above-mentioned image super-resolution system, further includes: the second construction unit, the second construction unit packet
It includes:
Subelement is obtained, for obtaining in convolutional neural networks convolutional layer and appointing based on the first processing image as input
One layer of fisrt feature；
Subelement is mapped, for carrying out mapping processing to the fisrt feature according to matrix recombination method, obtains the second spy
Sign；
Sample variance is calculated for the transposition based on the fisrt feature and second feature in feature calculation subelement
Matrix；
Matrix normalization subelement obtains covariance matrix for the sample variance matrix to be normalized；
Matrix computation subunit, for be based on the covariance matrix, be calculated the row mean value based on depth dimension to
Amount；
Learn subelement, is learned for successively carrying out dimensionality reduction study to the row mean vector based on depth dimension and rising dimension
It practises, obtains the first weight；
Weight normalizes subelement and obtains the second weight for first weight to be normalized；
Characteristic weighing subelement obtains weighted feature for being based on the fisrt feature and the second weight；
Subelement is constructed, for paying attention to power module using weighted feature building second order channel.
A kind of image super-resolution method and system provided based on the embodiments of the present invention, using image to be processed as in advance
The input of the convolutional neural networks oversubscription model first constructed, the convolutional neural networks oversubscription model sequentially connected are held by four
Row module composition, the second execution module are paid attention to the residual error module composition of power module by laminated structure and insertion second order channel；Via
The first execution module in the convolutional neural networks oversubscription model handles the image to be processed, obtains the first processing
Input of the image as second execution module；Feature is carried out to the first processing image using second execution module
Extract and characteristic processing, second processing image of the output comprising weighted feature as the convolutional neural networks oversubscription model the
The input of three execution modules；Size based on the third execution module and pre-set input picture and output image
Preset ratio coefficient carries out up-sampling treatment to the second processing image, obtains the third that size meets preset ratio coefficient
Image is handled, using third processing image as the defeated of the 4th execution module in the convolutional neural networks oversubscription model
Enter；Mapping processing, the corresponding image to be processed of output are carried out to third processing image via the 4th execution module
Super-resolution image.Based on the embodiment of the present invention, the convolutional neural networks oversubscription model is image to be processed setting weighting
Feature determines the important feature in the image to be processed by the study to the weighted feature, and according to important feature into
Row superresolution processing, so that the feature representation ability of the convolutional neural networks oversubscription model is improved, so that after superresolution processing
The details quality of obtained super-resolution image greatly improves.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
The embodiment of invention for those of ordinary skill in the art without creative efforts, can also basis
The attached drawing of offer obtains other attached drawings.
Fig. 1 is a kind of flow diagram of image super-resolution method provided in an embodiment of the present invention；
Fig. 2 is a kind of process signal of method for constructing convolutional neural networks oversubscription model provided in an embodiment of the present invention
Figure；
Fig. 3 is a kind of flow diagram of method for constructing the second execution module provided in an embodiment of the present invention；
Fig. 4 is a kind of structural schematic diagram of the residual error module with weighted feature provided in an embodiment of the present invention；
Fig. 5 is a kind of flow diagram for constructing second order channel attention modular approach provided in an embodiment of the present invention；
Fig. 6 is a kind of structural schematic diagram of image super-resolution system provided in an embodiment of the present invention；
Fig. 7 is the structural schematic diagram of another image super-resolution system provided in an embodiment of the present invention；
Fig. 8 is the structural schematic diagram of another image super-resolution system provided in an embodiment of the present invention；
Fig. 9 is the structural schematic diagram of another image super-resolution system provided in an embodiment of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
In this application, the terms "include", "comprise" or any other variant thereof is intended to cover non-exclusive inclusion,
So that the process, method, article or equipment for including a series of elements not only includes those elements, but also including not having
The other element being expressly recited, or further include for elements inherent to such a process, method, article, or device.Do not having
There is the element limited in the case where more limiting by sentence "including a ...", it is not excluded that in the mistake including the element
There is also other identical elements in journey, method, article or equipment.
As shown in Figure 1, being a kind of flow diagram of image super-resolution method provided in an embodiment of the present invention, the method
Include the following steps:
S101: using image to be processed as the input of the convolutional neural networks oversubscription model constructed in advance.
In S101, the convolutional neural networks oversubscription model is made of four sequentially connected execution modules, described to hold
Row module includes the first execution module, the second execution module, third execution module and the 4th execution module.Wherein, described second
Execution module is paid attention to the residual error module composition of power module by laminated structure and insertion second order channel.
It is inputted in the convolutional neural networks oversubscription model it should be noted that the second order channel notices that power module utilizes
Second-order statistics possessed by the feature of image, so that the weight of the convolutional neural networks oversubscription model adaptation learning characteristic
The property wanted, better focus utilization useful feature improve the feature representation ability of convolutional neural networks, so that it is super to improve image
The treatment effect of resolution.
S102: image to be processed is handled via the first execution module in convolutional neural networks oversubscription model, is obtained
Input to the first processing image as the second execution module.
In S102, the size of the first processing image is identical as the size of the image to be processed.
It should be noted that first execution module includes a convolutional layer, first execution module is based on the volume
Lamination carries out convolution algorithm to the image to be processed, obtains the first processing image.
S103: feature extraction and characteristic processing are carried out to the first processing image using the second execution module, output is comprising adding
Weigh input of the second processing image of feature as the third execution module in convolutional neural networks oversubscription model.
In S103, second execution module all multiple there is second order channel to pay attention to power module by itself
Residual error module, it is multi-time weighted to feature progress corresponding to the first processing image, so that it is determined that the first processing image
In feature of high importance, and obtain the second processing image comprising weighted feature.
It should be noted that because the second processing image includes weighted feature, in subsequent third execution module and the
In four execution module treatment processes, by the study to the weighted feature, the important feature in the image to be processed is determined,
And superresolution processing is carried out according to important feature, to improve the feature representation ability of the convolutional neural networks oversubscription model.
It needs to illustrate to be, the size of the second processing image is identical as the size of the image to be processed.
S104: the size preset ratio based on third execution module and pre-set input picture and output image
Coefficient handles second processing image, obtains the third processing image that size meets preset ratio coefficient, third is handled
Input of the image as the 4th execution module in convolutional neural networks oversubscription model.
In S104, the second processing image is input in the third execution module, the third execution module
The second processing image is amplified according to preset ratio coefficient, and is exported at the third to the 4th execution module
Manage image.
It should be noted that the third execution module includes a convolutional layer, the third execution module is based on the volume
Lamination carries out feature amplification to the second processing image, obtains the third processing image that size meets preset ratio coefficient.
It should be noted that the specific value of the preset ratio coefficient can be set according to the actual situation by technical staff
It sets.
S105: carrying out mapping processing to third processing image via the 4th execution module, the corresponding image to be processed of output
Super-resolution image.
In S105, the size of the super-resolution image is identical as the size of the image to be processed.
It should be noted that the 4th execution module includes a convolutional layer, the 4th execution module is based on the volume
Lamination maps third processing image, obtains and export the super-resolution image.
In embodiments of the present invention, using image to be processed as the defeated of the convolutional neural networks oversubscription model constructed in advance
Enter, the convolutional neural networks oversubscription model is made of four sequentially connected execution modules, and the second execution module is by sequentially folding
Add and be embedded in the residual error module composition that second order channel pays attention to power module；Via first in the convolutional neural networks oversubscription model
Execution module handles the image to be processed, obtains input of the first processing image as second execution module；
Feature extraction and characteristic processing are carried out to the first processing image using second execution module, output includes weighted feature
Second processing image as the convolutional neural networks oversubscription model third execution module input；It is held based on the third
The size preset ratio coefficient of row module and pre-set input picture and output image, to the second processing image
It is handled, the third processing image that size meets preset ratio coefficient is obtained, using third processing image as the volume
The input of the 4th execution module in product neural network oversubscription model；Third processing is schemed via the 4th execution module
As carrying out mapping processing, the super-resolution image of the corresponding image to be processed of output.Based on the embodiment of the present invention, the convolution
Neural network oversubscription model is that weighted feature is arranged in image to be processed, by the study to the weighted feature, determine it is described to
The important feature in image is handled, and carries out superresolution processing according to important feature, so that it is super to improve the convolutional neural networks
The feature representation ability of sub-model, so that the details quality of obtained super-resolution image greatly improves after superresolution processing.
Preferably, above-mentioned convolutional neural networks oversubscription model shown in fig. 1, in specific building process, with reference to Fig. 2, for this
A kind of flow diagram of the method for building convolutional neural networks oversubscription model that inventive embodiments provide, the method includes such as
Lower step:
S201: building training set.
In S201, the training set includes low-resolution image and high score corresponding with the low-resolution image
Resolution image.
S202: low-resolution image is input in preset convolutional neural networks model and carries out feature extraction, feature is put
Big and Feature Mapping, the image that obtains that treated.
In S202, in the preset convolutional neural networks model, convolution fortune is carried out to the low-resolution image
It calculates；Then, feature extraction is carried out to the low-resolution image after convolution algorithm by convolutional neural networks structure；Secondly,
Processing is amplified to the feature in the low-resolution image by up-sampling model；Finally, being based on and the low resolution
The corresponding high-definition picture of image carries out mapping processing to the amplified feature in the low resolution image, obtains described
Treated image.
S203: based on low-resolution image, full resolution pricture corresponding with low-resolution image and treated scheme
Picture, using preset loss function and the preset convolutional neural networks model of optimization algorithm training, until preset convolutional Neural
Network model exports high-definition picture corresponding with low-resolution image, determines the convolutional neural networks mould that current training obtains
Type is convolutional neural networks oversubscription model.
In S203, the convolutional neural networks oversubscription model is made of four sequentially connected execution modules, wherein institute
State the residual error module composition that the second execution module pays attention to power module by being superimposed and being embedded in second order channel.
It should be noted that in embodiments of the present invention, the loss function includes but is not limited to L1- norm, described excellent
Changing algorithm includes but is not limited to stochastic gradient descent algorithm.
In embodiments of the present invention, by construct training set, the training set include low-resolution image and with it is described
The corresponding high-definition picture of low-resolution image；By low-resolution image be input in preset convolutional neural networks model into
Row feature extraction, feature amplification and Feature Mapping, the image that obtains that treated；Based on low-resolution image and low resolution
The corresponding high-definition picture of image and treated image, it is default using preset loss function and optimization algorithm training
Convolutional neural networks model, until preset convolutional neural networks model exports corresponding with low-resolution image high score
Resolution image determines that the convolutional neural networks model that current training obtains is convolutional neural networks oversubscription model.Based on the present invention
Embodiment can effectively construct convolutional neural networks oversubscription model.
Preferably, above-mentioned shown in fig. 1, second execution module pays attention to power module by being superimposed and being embedded in second order channel
Residual error module composition, in specific building process, with reference to Fig. 3, for a kind of the second execution of building mould provided in an embodiment of the present invention
The flow diagram of the method for block, described method includes following steps:
S301: by preset second order channel attention Module-embedding into residual error module, obtain that there is the residual of weighted feature
Difference module.
In S301, the residual error module includes two convolutional layers, an active coating and two residual units.
Preferably, successively connect in sequence first convolutional layer, active coating, the second convolutional layer, the first residual unit,
Second order channel pays attention to power module and the second residual unit, obtains the residual error module with weighted feature.
In the concrete realization, the specific structure of the residual error module with weighted feature can refer to Fig. 4.
S302: the number of the residual error module with weighted feature needed for determining the second execution module of building.
In S302, the number of the residual error module with weighted feature is more, then the place of second execution module
It is better to manage effect.
It should be noted that the number of the residual error module needed for second execution module with weighted feature can be by skill
Art personnel are configured according to the actual situation, and the embodiment of the present invention is without limitation.
S303: each residual error module with weighted feature is stacked gradually, the second execution module is obtained.
In S303, convolutional neural networks structures can be constructed by stacking multiple residual error modules, in other words, described the
Two execution modules are the convolutional neural networks structure.Because convolutional neural networks structure has weighted feature as described in each
Residual error module stack forms, so weighted feature can be arranged for input picture in the convolutional neural networks structure.It follows that
Second execution module can generate the second processing image comprising weighted feature based on the first processing image.
In embodiments of the present invention, added by into residual error module, obtaining second order channel attention Module-embedding to have
Weigh the residual error module of feature；The number of residual error module with weighted feature needed for determining the second execution module of building；Successively
Each residual error module with weighted feature is stacked, second execution module is obtained.It, can based on the embodiment of the present invention
Effectively building obtains second execution module with second order channel attention mechanism.
Preferably, in the S301 shown in above-mentioned Fig. 3, the preset second order channel pays attention to this process of power module specific
In realization, with reference to Fig. 5, for a kind of process signal for constructing second order channel attention modular approach provided in an embodiment of the present invention
Figure, described method includes following steps:
S501: based on the first processing image as input, first of any layer in convolutional neural networks convolutional layer is obtained
Feature.
In S501, the fisrt feature is specially H × W × C characteristic pattern, and by the H × W × C feature icon
It is denoted as x, wherein H is the height of convolutional layer, and W is the width of convolutional layer, and C is the depth of convolutional layer.
S502: mapping processing is carried out to fisrt feature according to matrix recombination method, obtains second feature.
In S502, according to the matrix recombination method, the H × W × C characteristic pattern x is mapped as (H*W) × C's
Feature X, and X is exported, X is the second feature.
S503: sample variance matrix is calculated in the transposition based on fisrt feature and second feature.
In S503, sample variance matrix Σ is calculated according to formula (1).
Wherein,I refers to the unit matrix of (H*W) × (H*W), and 1 refers to
Element is all 1 matrix.
S504: being normalized sample variance matrix, obtains covariance matrix.
In S504, according to formula (2), sample variance matrix Σ is normalized, covariance matrix Y is obtained.
Y=Σ0.5=U Λ0.5UT (2)
Wherein, Λ=diag (λ1,...,λC), U refers to that orthogonal matrix, Λ refer to that element is characterized value λiIt is diagonal
Matrix, each eigenvalue λ in the diagonal matrixiIt is ranked up according to the sequence of descending, λ is positive integer, and i refers to described
The columns of diagonal matrix.
S505: it is based on covariance matrix, the row mean vector based on depth dimension is calculated.
In S505, the jth row element mean value of covariance matrix Y is calculated according to formula (3), and according to the jth
Row element mean value determines the row mean vector of C dimension.
Wherein, j refers to the line number of the diagonal matrix.
S506: successively carrying out dimensionality reduction study to the row mean vector based on depth dimension and rises dimension study, obtains the first power
Weight.
In S506, using the row mean vector of C dimension as the input of preset first fully-connected network, to 1 × 1 × C/r
It is exported, obtains output result m.Using the output result m as the input of preset second fully-connected network, to 1 × 1 ×
C is exported, and first weight is obtained.
S507: being normalized the first weight, obtains the second weight.
In S507, calculating is normalized to first weight using Sigmoid function, obtains second power
Weight.
S508: it is based on fisrt feature and the second weight, obtains weighted feature.
In S508, the fisrt feature is multiplied with the second weight, obtains the weighted feature.
S509: power module is paid attention to using weighted feature building second order channel.
In S509, the weighted feature is embedded into convolutional neural networks, the second order channel is generated and pays attention to power module.
In embodiments of the present invention, it based on the first processing image as input, obtains in convolutional neural networks convolutional layer
The fisrt feature of any layer；Mapping processing is carried out to the fisrt feature according to matrix recombination method, obtains second feature；It is based on
The fisrt feature and second feature, are calculated sample variance matrix；The sample variance matrix is normalized,
Obtain covariance matrix；Based on the covariance matrix, the row mean vector based on depth dimension is calculated；It is based on to described
The row mean vector of depth dimension carries out dimensionality reduction study and rises dimension study, obtains the first weight；First weight is returned
One change processing, obtains the second weight；Based on the fisrt feature and the second weight, weighted feature is obtained, and utilizes the weighting
Feature construction second order channel pays attention to power module.Based on the embodiment of the present invention, it can effectively construct second order channel and pay attention to power module.
Based on a kind of image super-resolution method that the embodiments of the present invention provide, also correspondence of the embodiment of the present invention is provided
A kind of image super-resolution system, as shown in fig. 6, being a kind of structural representation of image super-resolution system provided in an embodiment of the present invention
Figure, the system comprises:
Input unit 100, for using image to be processed as the input of the convolutional neural networks oversubscription model constructed in advance,
The convolutional neural networks oversubscription model is made of four sequentially connected execution modules, and the second execution module is by being superimposed and being embedded in
Second order channel pays attention to the residual error module composition of power module.
First execution unit 200, for via the first execution module in the convolutional neural networks oversubscription model to institute
It states image to be processed to be handled, obtains input of the first processing image as second execution module, first processing
The size of image is identical as the size of the image to be processed.
Second execution unit 300 is mentioned for carrying out feature to the first processing image using second execution module
It takes and characteristic processing, second processing image of the output comprising weighted feature is as the in the convolutional neural networks oversubscription model
The input of three execution modules.
Third execution unit 400, for based on the third execution module and pre-set input picture and output
The size preset ratio coefficient of image, handles the second processing image, obtains size and meets preset ratio coefficient
Third handles image, using third processing image as the 4th execution module in the convolutional neural networks oversubscription model
Input.
4th execution unit 500, it is defeated for carrying out mapping processing to third processing image via the 4th execution module
The super-resolution image of the image to be processed is corresponded to out.
In embodiments of the present invention, using image to be processed as the defeated of the convolutional neural networks oversubscription model constructed in advance
Enter, the convolutional neural networks oversubscription model is made of four sequentially connected execution modules, and the second execution module is by sequentially folding
Add and be embedded in the residual error module composition that second order channel pays attention to power module；Via first in the convolutional neural networks oversubscription model
Execution module handles the image to be processed, obtains input of the first processing image as second execution module；
Feature extraction and characteristic processing are carried out to the first processing image using second execution module, output includes weighted feature
Second processing image as the convolutional neural networks oversubscription model third execution module input；It is held based on the third
The size preset ratio coefficient of row module and pre-set input picture and output image, to the second processing image
It is handled, the third processing image that size meets preset ratio coefficient is obtained, using third processing image as the volume
The input of the 4th execution module in product neural network oversubscription model；Third processing is schemed via the 4th execution module
As carrying out mapping processing, the super-resolution image of the corresponding image to be processed of output.Based on the embodiment of the present invention, the convolution
Neural network oversubscription model is that weighted feature is arranged in image to be processed, by the study to the weighted feature, determine it is described to
The important feature in image is handled, and carries out superresolution processing according to important feature, so that it is super to improve the convolutional neural networks
The feature representation ability of sub-model, so that the details quality of obtained super-resolution image greatly improves after superresolution processing.
Preferably, in conjunction with Fig. 6, with reference to Fig. 7, for the structure of another image super-resolution system provided in an embodiment of the present invention
Schematic diagram, the input unit 100 include:
Construct subelement 101, for constructing training set, the training set include low-resolution image and with it is described low
The corresponding high-definition picture of image in different resolution.
Handle subelement 102, for by the low-resolution image be input in preset convolutional neural networks model into
Row feature extraction, feature amplification and Feature Mapping, the image that obtains that treated.
Training subelement 103, for being based on the low-resolution image, high-resolution corresponding with the low-resolution image
Rate image and treated image utilize preset loss function and the optimization algorithm training preset convolutional Neural net
Network model, until the preset convolutional neural networks model exports high resolution graphics corresponding with the low-resolution image
Picture determines that the convolutional neural networks model that current training obtains is convolutional neural networks oversubscription model；Wherein, the convolutional Neural
Network oversubscription model is made of four sequentially connected execution modules, and the second execution module is paid attention to by being superimposed and being embedded in second order channel
The residual error module composition of power module.
In embodiments of the present invention, by construct training set, the training set include low-resolution image and with it is described
The corresponding high-definition picture of low-resolution image；By low-resolution image be input in preset convolutional neural networks model into
Row feature extraction, feature amplification and Feature Mapping, the image that obtains that treated；Based on low-resolution image and low resolution
The corresponding high-definition picture of image and treated image, it is default using preset loss function and optimization algorithm training
Convolutional neural networks model, until preset convolutional neural networks model exports corresponding with low-resolution image high score
Resolution image determines that the convolutional neural networks model that current training obtains is convolutional neural networks oversubscription model.Based on the present invention
Embodiment can effectively construct convolutional neural networks oversubscription model.
Preferably, in conjunction with Fig. 6, with reference to Fig. 8, for the structure of another image super-resolution system provided in an embodiment of the present invention
Schematic diagram, the system also includes:
First construction unit 600, for second order channel attention Module-embedding into residual error module, to be obtained to have weighting
The residual error module of feature, determining has the number of the residual error module of weighted feature needed for the second execution module of building, successively heap
Each residual error module with weighted feature is folded, second execution module is obtained.
Preferably, first construction unit 600 be specifically used for successively connect in sequence the first convolutional layer, active coating,
Second convolutional layer, second order channel pay attention to power module and residual error, obtain the residual error module with weighted feature.
In embodiments of the present invention, added by into residual error module, obtaining second order channel attention Module-embedding to have
Weigh the residual error module of feature；The number of residual error module with weighted feature needed for determining the second execution module of building；Successively
Each residual error module with weighted feature is stacked, second execution module is obtained.It, can based on the embodiment of the present invention
Effectively building obtains second execution module with second order channel attention mechanism.
Preferably, in conjunction with Fig. 8, with reference to Fig. 9, for the structure of another image super-resolution system provided in an embodiment of the present invention
Schematic diagram, the system also includes: the second construction unit 700.
Second construction unit 700 includes:
Subelement 701 is obtained, for obtaining in convolutional neural networks convolutional layer based on the first processing image as input
The fisrt feature of any layer.
Subelement 702 is mapped, for carrying out mapping processing to the fisrt feature according to matrix recombination method, obtains second
Feature.
Sample side is calculated for the transposition based on the fisrt feature and second feature in feature calculation subelement 703
Poor matrix.
Matrix normalization subelement 704 obtains covariance square for the sample variance matrix to be normalized
Battle array.
The row mean value based on depth dimension is calculated for being based on the covariance matrix in matrix computation subunit 705
Vector.
Learn subelement 706, for successively carrying out dimensionality reduction study and liter to the row mean vector based on depth dimension
Dimension study, obtains the first weight.
Weight normalizes subelement 707 and obtains the second weight for first weight to be normalized.
Characteristic weighing subelement 708 obtains weighted feature for being based on the fisrt feature and the second weight.
Subelement 709 is constructed, for paying attention to power module using weighted feature building second order channel.
In embodiments of the present invention, it based on the first processing image as input, obtains in convolutional neural networks convolutional layer
The fisrt feature of any layer；Mapping processing is carried out to the fisrt feature according to matrix recombination method, obtains second feature；It is based on
The transposition of the fisrt feature and second feature, is calculated sample variance matrix；Normalizing is carried out to the sample variance matrix
Change processing, obtains covariance matrix；Based on the covariance matrix, the row mean vector based on depth dimension is calculated；It is right
The row mean vector based on depth dimension carries out dimensionality reduction study and rises dimension study, obtains the first weight；To first power
It is normalized again, obtains the second weight；Based on the fisrt feature and the second weight, weighted feature is obtained, and is utilized
Weighted feature building second order channel pays attention to power module.Based on the embodiment of the present invention, it can effectively construct second order channel and pay attention to
Power module.
All the embodiments in this specification are described in a progressive manner, same and similar portion between each embodiment
Dividing may refer to each other, and each embodiment focuses on the differences from other embodiments.Especially for system or
For system embodiment, since it is substantially similar to the method embodiment, so describing fairly simple, related place is referring to method
The part of embodiment illustrates.System and system embodiment described above is only schematical, wherein the conduct
The unit of separate part description may or may not be physically separated, component shown as a unit can be or
Person may not be physical unit, it can and it is in one place, or may be distributed on multiple network model units.It can
It is achieved the purpose of the solution of this embodiment with selecting some or all of the modules therein according to the actual needs.This field is common
Technical staff can understand and implement without creative efforts.
Professional further appreciates that, unit described in conjunction with the examples disclosed in the embodiments of the present disclosure
And algorithm steps, can be realized with electronic hardware, computer software, or a combination of the two, in order to clearly demonstrate hardware and
The interchangeability of software generally describes each exemplary composition and step according to function in the above description.These
Function is implemented in hardware or software actually, the specific application and design constraint depending on technical solution.Profession
Technical staff can use different methods to achieve the described function each specific application, but this realization is not answered
Think beyond the scope of this invention.
The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention.
Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, of the invention
It is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein phase one
The widest scope of cause.
Claims (10)
1. a kind of image super-resolution method, which is characterized in that the described method includes:
Using image to be processed as the input of the convolutional neural networks oversubscription model constructed in advance, the convolutional neural networks oversubscription
Model is made of four sequentially connected execution modules, and the second execution module pays attention to power module by being superimposed and being embedded in second order channel
Residual error module composition；
The image to be processed is handled via the first execution module in the convolutional neural networks oversubscription model, is obtained
Input of the first processing image as second execution module, the size of the first processing image and the image to be processed
Size it is identical；
Feature extraction and characteristic processing are carried out to the first processing image using second execution module, output includes weighting
Input of the second processing image of feature as the third execution module in the convolutional neural networks oversubscription model；
Size preset ratio coefficient based on the third execution module and pre-set input picture and output image,
The second processing image is handled, the third processing image that size meets preset ratio coefficient is obtained, by the third
Handle input of the image as the 4th execution module in the convolutional neural networks oversubscription model；
Mapping processing is carried out to third processing image via the 4th execution module, the corresponding image to be processed of output surpasses
Image in different resolution.
2. the method according to claim 1, wherein the convolutional neural networks oversubscription model constructed in advance
Process, comprising:
Training set is constructed, the training set includes low-resolution image and high-resolution corresponding with the low-resolution image
Rate image；
By the low-resolution image be input in preset convolutional neural networks model carry out feature extraction, feature amplification and
Feature Mapping, the image that obtains that treated；
Based on the low-resolution image, high-definition picture corresponding with the low-resolution image and treated figure
Picture, using preset loss function and the optimization algorithm training preset convolutional neural networks model, until described preset
Convolutional neural networks model exports high-definition picture corresponding with the low-resolution image, determines the volume that current training obtains
Product neural network model is convolutional neural networks oversubscription model；
Wherein, the convolutional neural networks oversubscription model is made of four sequentially connected execution modules, the second execution module by
Superposition and insertion second order channel pay attention to the residual error module composition of power module.
3. the method according to claim 1, wherein it is described superposition and insertion second order channel pay attention to the residual of power module
Difference module constitutes the process of the second execution module, comprising:
By preset second order channel attention Module-embedding into residual error module, the residual error module with weighted feature is obtained；
The number of residual error module with weighted feature needed for determining the second execution module of building；
Each residual error module with weighted feature is stacked gradually, second execution module is obtained.
4. according to the method described in claim 3, it is characterized in that, described by second order channel attention Module-embedding to residual error mould
In block, the residual error module with weighted feature is obtained, comprising:
The first convolutional layer, active coating, the second convolutional layer, the first residual unit, second order channel attention are successively connected in sequence
Module and the second residual unit obtain the residual error module with weighted feature.
5. according to the method described in claim 4, it is characterized in that, the default second order channel pays attention to the process of power module, packet
It includes:
Based on the first processing image as input, the fisrt feature of any layer in convolutional neural networks convolutional layer is obtained；
Mapping processing is carried out to the fisrt feature according to matrix recombination method, obtains second feature；
Transposition based on the fisrt feature and second feature, is calculated sample variance matrix；
The sample variance matrix is normalized, covariance matrix is obtained；
Based on the covariance matrix, the row mean vector based on depth dimension is calculated；
Dimensionality reduction study is successively carried out to the row mean vector based on depth dimension and rises dimension study, obtains the first weight；
First weight is normalized, the second weight is obtained；
Based on the fisrt feature and the second weight, weighted feature is obtained；
Power module is paid attention to using weighted feature building second order channel.
6. a kind of image super-resolution system characterized by comprising
Input unit, for using image to be processed as the input of the convolutional neural networks oversubscription model constructed in advance, the volume
Product neural network oversubscription model is made of four sequentially connected execution modules, and the second execution module is led to by being superimposed and being embedded in second order
Road pays attention to the residual error module composition of power module；
First execution unit, for via the first execution module in the convolutional neural networks oversubscription model to described to be processed
Image is handled, and input of the first processing image as second execution module, the ruler of the first processing image are obtained
It is very little identical as the size of the image to be processed；
Second execution unit, for carrying out feature extraction and feature to the first processing image using second execution module
Processing, second processing image of the output comprising weighted feature execute mould as the third in the convolutional neural networks oversubscription model
The input of block；
Third execution unit, for based on the third execution module and pre-set input picture and output image
Size preset ratio coefficient handles the second processing image, obtains size and meets at the third of preset ratio coefficient
Image is managed, using third processing image as the input of the 4th execution module in the convolutional neural networks oversubscription model；
4th execution unit, for carrying out mapping processing to third processing image via the 4th execution module, output is corresponded to
The super-resolution image of the image to be processed.
7. system according to claim 6, which is characterized in that the input unit includes:
Construct subelement, for constructing training set, the training set include low-resolution image and with the low resolution figure
As corresponding high-definition picture；
Subelement is handled, is mentioned for the low-resolution image to be input to progress feature in preset convolutional neural networks model
It takes, feature amplification and Feature Mapping, the image that obtains that treated；
Training subelement, for based on the low-resolution image, high-definition picture corresponding with the low-resolution image,
And treated image, using preset loss function and the optimization algorithm training preset convolutional neural networks model,
Until the preset convolutional neural networks model exports high-definition picture corresponding with the low-resolution image, determination is worked as
The convolutional neural networks model that preceding training obtains is convolutional neural networks oversubscription model；Wherein, the convolutional neural networks oversubscription
Model is made of four sequentially connected execution modules, and the second execution module pays attention to power module by being superimposed and being embedded in second order channel
Residual error module composition.
8. system according to claim 6, which is characterized in that further include:
First construction unit, for preset second order channel attention Module-embedding into residual error module, to be obtained to have weighting
The residual error module of feature, determining has the number of the residual error module of weighted feature needed for the second execution module of building, successively heap
Each residual error module with weighted feature is folded, second execution module is obtained.
9. system according to claim 8, which is characterized in that described by second order channel attention Module-embedding to residual error mould
In block, the first construction unit for obtaining the residual error module with weighted feature is specifically used for successively connecting the first convolution in sequence
Layer, active coating, the second convolutional layer, second order channel pay attention to power module and residual error, obtain the residual error module with weighted feature.
10. system according to claim 9, which is characterized in that further include: the second construction unit, second building are single
Member includes:
Subelement is obtained, for obtaining any layer in convolutional neural networks convolutional layer based on the first processing image as input
Fisrt feature；
Subelement is mapped, for carrying out mapping processing to the fisrt feature according to matrix recombination method, obtains second feature；
Sample variance matrix is calculated for the transposition based on the fisrt feature and second feature in feature calculation subelement；
Matrix normalization subelement obtains covariance matrix for the sample variance matrix to be normalized；
The row mean vector based on depth dimension is calculated for being based on the covariance matrix in matrix computation subunit；
Learn subelement, for successively carrying out dimensionality reduction study to the row mean vector based on depth dimension and rising dimension study,
Obtain the first weight；
Weight normalizes subelement and obtains the second weight for first weight to be normalized；
Characteristic weighing subelement obtains weighted feature for being based on the fisrt feature and the second weight；
Subelement is constructed, for paying attention to power module using weighted feature building second order channel.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910439532.8A CN110175953B (en) | 2019-05-24 | 2019-05-24 | Image super-resolution method and system |
| PCT/CN2020/088215 WO2020238558A1 (en) | 2019-05-24 | 2020-04-30 | Image super-resolution method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910439532.8A CN110175953B (en) | 2019-05-24 | 2019-05-24 | Image super-resolution method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110175953A true CN110175953A (en) | 2019-08-27 |
| CN110175953B CN110175953B (en) | 2023-04-18 |
Family
ID=67692105
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910439532.8A Active CN110175953B (en) | 2019-05-24 | 2019-05-24 | Image super-resolution method and system |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN110175953B (en) |
| WO (1) | WO2020238558A1 (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020238558A1 (en) * | 2019-05-24 | 2020-12-03 | 鹏城实验室 | Image super-resolution method and system |
| CN112767427A (en) * | 2021-01-19 | 2021-05-07 | 西安邮电大学 | Low-resolution image recognition algorithm for compensating edge information |
| CN116775938A (en) * | 2023-08-15 | 2023-09-19 | 腾讯科技（深圳）有限公司 | Method, device, electronic equipment and storage medium for retrieving comment video |
Families Citing this family (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112634238B (en) * | 2020-12-25 | 2024-03-08 | 武汉大学 | Attention module-based image quality evaluation method |
| CN112991173B (en) * | 2021-03-12 | 2024-04-16 | 西安电子科技大学 | Single-frame image super-resolution reconstruction method based on dual-channel feature migration network |
| CN113538231B (en) * | 2021-06-17 | 2024-04-02 | 杭州电子科技大学 | Single image super-resolution reconstruction system and method based on pixel distribution estimation |
| CN113538616B (en) * | 2021-07-09 | 2023-08-18 | 浙江理工大学 | Magnetic resonance image reconstruction method combining PUGAN with improved U-net |
| CN113610706A (en) * | 2021-07-19 | 2021-11-05 | 河南大学 | Fuzzy monitoring image super-resolution reconstruction method based on convolutional neural network |
| CN113538244B (en) * | 2021-07-23 | 2023-09-01 | 西安电子科技大学 | Lightweight super-resolution reconstruction method based on self-adaptive weight learning |
| CN113837941B (en) * | 2021-09-24 | 2023-09-01 | 北京奇艺世纪科技有限公司 | Training method and device for image superdivision model and computer readable storage medium |
| CN113706388B (en) * | 2021-09-24 | 2023-06-27 | 上海壁仞智能科技有限公司 | Image super-resolution reconstruction method and device |
| CN114022355B (en) * | 2021-09-26 | 2024-02-20 | 陕西师范大学 | Image super-resolution method based on recursive attention mechanism |
| CN115082307B (en) * | 2022-05-14 | 2024-03-19 | 西北工业大学深圳研究院 | Image super-resolution method based on fractional differential equation |
| CN115100042A (en) * | 2022-07-20 | 2022-09-23 | 北京工商大学 | Pathological image super-resolution method based on channel attention retention network |
| CN115564649B (en) * | 2022-09-27 | 2023-07-28 | 苏州大学 | Image super-resolution reconstruction method, device and equipment |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2007316161A (en) * | 2006-05-23 | 2007-12-06 | Matsushita Electric Ind Co Ltd | Super resolution processing method and apparatus using residual interpolation |
| US20180075581A1 (en) * | 2016-09-15 | 2018-03-15 | Twitter, Inc. | Super resolution using a generative adversarial network |
| CN108734660A (en) * | 2018-05-25 | 2018-11-02 | 上海通途半导体科技有限公司 | A kind of image super-resolution rebuilding method and device based on deep learning |
| CN108765296A (en) * | 2018-06-12 | 2018-11-06 | 桂林电子科技大学 | A kind of image super-resolution rebuilding method based on recurrence residual error attention network |
| CN108921786A (en) * | 2018-06-14 | 2018-11-30 | 天津大学 | Image super-resolution reconstructing method based on residual error convolutional neural networks |
| CN108960261A (en) * | 2018-07-25 | 2018-12-07 | 扬州万方电子技术有限责任公司 | A kind of obvious object detection method based on attention mechanism |
| CN109741260A (en) * | 2018-12-29 | 2019-05-10 | 天津大学 | A kind of efficient super-resolution method based on depth back projection network |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109584161A (en) * | 2018-11-29 | 2019-04-05 | 四川大学 | The Remote sensed image super-resolution reconstruction method of convolutional neural networks based on channel attention |
| CN110175953B (en) * | 2019-05-24 | 2023-04-18 | 鹏城实验室 | Image super-resolution method and system |
-
2019
- 2019-05-24 CN CN201910439532.8A patent/CN110175953B/en active Active
-
2020
- 2020-04-30 WO PCT/CN2020/088215 patent/WO2020238558A1/en active Application Filing
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2007316161A (en) * | 2006-05-23 | 2007-12-06 | Matsushita Electric Ind Co Ltd | Super resolution processing method and apparatus using residual interpolation |
| US20180075581A1 (en) * | 2016-09-15 | 2018-03-15 | Twitter, Inc. | Super resolution using a generative adversarial network |
| CN108734660A (en) * | 2018-05-25 | 2018-11-02 | 上海通途半导体科技有限公司 | A kind of image super-resolution rebuilding method and device based on deep learning |
| CN108765296A (en) * | 2018-06-12 | 2018-11-06 | 桂林电子科技大学 | A kind of image super-resolution rebuilding method based on recurrence residual error attention network |
| CN108921786A (en) * | 2018-06-14 | 2018-11-30 | 天津大学 | Image super-resolution reconstructing method based on residual error convolutional neural networks |
| CN108960261A (en) * | 2018-07-25 | 2018-12-07 | 扬州万方电子技术有限责任公司 | A kind of obvious object detection method based on attention mechanism |
| CN109741260A (en) * | 2018-12-29 | 2019-05-10 | 天津大学 | A kind of efficient super-resolution method based on depth back projection network |
Non-Patent Citations (3)
| Title |
|---|
| YUE LU 等: "Channel Attention and Multi-level Features Fusion for Single Image Super-Resolution", 《2018 IEEE VISUAL COMMUNICATIONS AND IMAGE PROCESSING (VCIP)》 * |
| 徐诚极 等: "Attention-YOLO：引入注意力机制的YOLO检测算法", 《计算机工程与应用》 * |
| 王东飞: "基于通道注意力的卷积神经网络在图像超分辨率重建中的应用", 《广播与电视技术》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020238558A1 (en) * | 2019-05-24 | 2020-12-03 | 鹏城实验室 | Image super-resolution method and system |
| CN112767427A (en) * | 2021-01-19 | 2021-05-07 | 西安邮电大学 | Low-resolution image recognition algorithm for compensating edge information |
| CN116775938A (en) * | 2023-08-15 | 2023-09-19 | 腾讯科技（深圳）有限公司 | Method, device, electronic equipment and storage medium for retrieving comment video |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110175953B (en) | 2023-04-18 |
| WO2020238558A1 (en) | 2020-12-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110175953A (en) | A kind of image super-resolution method and system | |
| US10810483B2 (en) | Superpixel methods for convolutional neural networks | |
| CN105512723A (en) | Artificial neural network calculating device and method for sparse connection | |
| Diaconis et al. | Geometric bounds for eigenvalues of Markov chains | |
| CN108805270A (en) | A kind of convolutional neural networks system based on memory | |
| CN107533667A (en) | Vector calculation unit in neural network processor | |
| CN106959937B (en) | A kind of vectorization implementation method of the warp product matrix towards GPDSP | |
| CN107480774A (en) | Dynamic neural network model training method and device based on integrated study | |
| CN106682734A (en) | Method and apparatus for increasing generalization capability of convolutional neural network | |
| CN108596267A (en) | A kind of image rebuilding method, terminal device and computer readable storage medium | |
| CN108334945A (en) | The acceleration of deep neural network and compression method and device | |
| CN112529165B (en) | Deep neural network pruning method, device, terminal and storage medium | |
| CN107480788A (en) | A kind of training method and training system of depth convolution confrontation generation network | |
| CN109903152A (en) | Default loss analysis method, device and computer readable storage medium | |
| CN110222598A (en) | A kind of video behavior recognition methods, device, storage medium and server | |
| CN109992742A (en) | A kind of signal processing method and device | |
| Kirk et al. | Vassiliev invariants of two component links and the Casson-Walker invariant | |
| Smith | E6, Strings, Branes, and the Standard Model | |
| CN104573727B (en) | A kind of handwriting digital image dimension reduction method | |
| CN114360032B (en) | Polymorphic invariance face recognition method and system | |
| CN109558944A (en) | The algorithm optimization method and device of convolutional neural networks based on configurable convolutional layer | |
| Zachariah et al. | A multicriteria optimization model for quality of modular software systems | |
| CN107633010B (en) | Identification method and system for GRC plate image with complex modeling | |
| Mladenov et al. | Solving sudoku puzzles by using hopfield neural networks | |
| CN110033087A (en) | Based on neural network processor and its application method |
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 |