CN107358575A - A kind of single image super resolution ratio reconstruction method based on depth residual error network - Google Patents
A kind of single image super resolution ratio reconstruction method based on depth residual error network Download PDFInfo
- Publication number
- CN107358575A CN107358575A CN201710426614.XA CN201710426614A CN107358575A CN 107358575 A CN107358575 A CN 107358575A CN 201710426614 A CN201710426614 A CN 201710426614A CN 107358575 A CN107358575 A CN 107358575A
- Authority
- CN
- China
- Prior art keywords
- image
- resolution
- training
- residual error
- low
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4046—Scaling the whole image or part thereof using neural networks
-
- 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
- G06T3/4076—Super resolution, i.e. output image resolution higher than sensor resolution by iteratively correcting the provisional high resolution image using the original low-resolution image
Abstract
The invention discloses a kind of single image super resolution ratio reconstruction method based on depth residual error network.The invention mainly includes steps：1st, block extraction and pixel average treatment are carried out to the image in sample image database, obtains corresponding high-resolution and low resolution training image collection；2nd, a depth convolutional neural networks with residual error structure are built and are iterated training, and then the neutral net built in the training set input step two obtained in step 1 is iterated training；3rd, the data model obtained according to training, it is implemented in combination with amplifying the continuous ratio for inputting low-resolution image by interative computation and interpolation algorithm.The present invention is by introducing depth residual error network, simultaneously up-sampling layer is introduced in network end-point, accelerate the processing speed of image amplification, enhance the display effect of image detail part, better image super-resolution rebuilding effect is obtained, is shown in image high definition, compression of images, safety inspection etc. have a wide range of applications in field.
Description
Technical field
The invention belongs to technical field of image processing, and in particular to a kind of single image oversubscription based on depth residual error network
Resolution method for reconstructing.
Background technology
At present, with the extensive use of multimedia technology and a wide range of popularization of digital equipment, digital picture due to its into
This is low, real-time is good, is easy to the advantages such as post-processing to achieve extensive use in each field.However, but by sensor
Physical size sum purpose limits, and the spatial resolution of digital picture caused by imaging system is often difficult to meet check of drawings personnel's
Demand, and being limited by image-forming principle and manufacture craft, manufactures the imaging system of high spatial resolution and will be significantly increased and be
The cost of system and R＆D cycle.Therefore in hardware system under the same conditions, the resolution ratio of digital picture is lifted using algorithm, it is extensive
Multiple high-definition picture details is significant.
Image super-resolution rebuilding refers to obtain high-definition picture using low-resolution image or image sequence, according to weight
3 classes can be divided into by building thinking：Interpolation algorithm, algorithm for reconstructing and the method based on study.Wherein weighed using the pixel of neighborhood
Recombination obtains the interpolation algorithm of target pixel value because the characteristics of its form is simple, processing speed is fast is widely used in digitizing
In X-ray shooting system.However, this kind of algorithm can lose image high-frequency information, image detail is lost, causes image to obscure, no
Accurate judgement is made beneficial to check of drawings personnel.Occurs the interpolation algorithm based on edge on this basis afterwards, it is to a certain degree
On remain the high-frequency information of image, but the interpolation algorithm based on edge is difficult to the texture region for handling image, the scope of application
There is very big limitation.
Different from traditional super-resolution method, the method based on study establishes low resolution using the machine learning for having supervision
The Nonlinear Mapping relation of image and high-definition picture carries out the reconstruction of image.This kind of method can extract image from sample set
Prior information, so as to obtain higher reconstruction precision.However, traditional method for reconstructing based on study can only extract image
The single order of relatively simple feature, such as image, second order gradient, these features are not sufficient to adequately characterize image information, make figure
The reconstruction quality of picture is restricted.Inspired by the immense success that deep learning method obtains in computer vision field,
Deep learning method is applied to image super-resolution rebuilding problem, the single-frame images oversubscription using convolutional neural networks occurs
Resolution algorithm for reconstructing (SRCNN), artificial design feature extracting mode is avoided, end-to-end study is realized, improves image
Reconstruction precision.However, due to the influence of gradient diffusing phenomenon in convolutional neural networks, SRCNN is in the larger feelings of convolution depth
The phenomenon that network is degenerated occurs under condition, i.e. the reconstruction quality of image declines, and limits the reconstruction performance of algorithm.
Following defect be present in existing image super-resolution rebuilding method：1. being concentrated in extraction training sample, do not account for
The property of image in itself, random process is carried out to the high and low frequency part of image, training sample is may result in and concentrates low frequency
Partial information redundancy, limit neural metwork training effect；2. substantial treatment object is mostly bicubic interpolation figure in the prior art
Picture, and bicubic interpolation is substantially LPF in image procossing, may lose the high-frequency information of image, therefore on image side
The part such as edge, texture can not obtain preferably rebuilding effect, and carry out processing to interpolation image and can greatly increase reconstruction tasks
Amount of calculation；3. network can not solve the gradient diffusing phenomenon being likely to occur in neural metwork training in the prior art, network is deep
Degree is unable to reach ten layers, it is difficult to fully characterizes image information, limits the effect of image reconstruction.4. nerve net in the prior art
Network designs for the magnification ratio of a certain fixation, it is difficult to realizes more flexible continuous ratio amplification.
The content of the invention
Not high for some existing super resolution technology reconstruction precisions, it is slow to rebuild speed, can not continuous ratio enlarged drawing etc.
Shortcoming.The application proposes a kind of single image super resolution ratio reconstruction method based on depth residual error network.For existing method weight
The defects of precision is not high is built, residual error structure is introduced neutral net, eliminates the gradient diffusing phenomenon in neutral net by the present invention,
The depth of neutral net is considerably increased, improves the speed of training process and the precision of reconstruction image.For existing method weight
Slow-footed defect is built, the present invention introduces up-sampling layer in neutral net last layer, while reduces the parameter number in neutral net
Mesh, so as to improve the speed of image reconstruction.For existing method can not continuous ratio enlarged drawing the shortcomings that, the present invention scheming
In the process of reconstruction of picture, by low-resolution image by neutral net iterative processing, and using interpolation algorithm be sized from
And realize the continuous ratio amplification of image.
Brief description of the drawings
The step of Fig. 1 is carry out image reconstruction of the present invention based on depth residual error network；
Fig. 2 is the step of present invention obtains the file for rebuilding training；
Fig. 3 is the neural network structure of the ultra-deep level in the present invention；
Fig. 4 for the present invention in neutral net to image procossing the step of；
Fig. 5 is the Contrast on effect of the amplification of prior art and the present invention to image；
Fig. 6 is the index contrast of the amplification of prior art and the present invention to image.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as in addition, technical characteristic involved in each embodiment of invention described below
Conflict can is not formed each other to be mutually combined.
The single image super resolution ratio reconstruction method based on depth residual error network of the present invention, as shown in Figure 1 including following
Step：1st, block extraction and pixel average treatment are carried out to the image in sample image database, obtains corresponding high-resolution
With low resolution training image collection；2nd, a depth convolutional neural networks with residual error structure are built and are iterated training, are entered
And the neutral net built in the training set input step two obtained in step 1 is iterated training；3rd, obtained according to training
Data model, by interative computation and interpolation algorithm be implemented in combination with to input low-resolution image continuous ratio amplify.
The single image super resolution ratio reconstruction method based on depth residual error network of the present invention, it is necessary first to obtain being used for weight
The file of training is built, wherein being most importantly：It is image in the higher region of image gradient 1. considering the Gradient Features of image
HFS extraction is more intensive, and low frequency part extracts more sparse.It is more can so to ensure that the high-frequency region in training set accounts for
Number, help to lift the training effect of neutral net；2. in addition to extraction high-low resolution figure area, the figure of bicubic interpolation is added
Area, described bicubic interpolation image contain the low frequency part of image, it is ensured that neutral net is only carried out to high-frequency characteristic
Study, reduces training difficulty, enhances training effect.
In image scale operation, output image pixel point coordinates may correspond on input picture between several pixels
Position, it is necessary to calculate the gray value of the output point by gray-level interpolation processing.Bicubic interpolation is not only examined in the present invention
Consider the influence of four direct neighbor pixel gray values of surrounding, it is also contemplated that the influence of their gray-value variation rates.This
In method, function f point (x, y) value f (x, y) can by rectangular mesh in nearest 16 sampled points weighted average
Obtain, need to use two polynomial interopolation cubic functions herein, each direction uses one.Can be with by bicubic interpolation
A continuous interpolating function is obtained, its first-order partial derivative is continuous, and cross derivative everywhere continuous.Method of weighting is
F (i+u, j+v)=ABCT
A=[S (u+1) S (u) S (u-1) S (u-2)]
C=[S (v+1) S (v) S (v-1) S (v-2)]
B=f (i-1：I+2, j-1：j+2)
Wherein S (x) is cubic interpolation kernel function, can be approximate by following formula：
Bicubic interpolation is substantially the low pass filter of image in the present invention, can obtain smoother image, but
The high-frequency information of image is filtered off in interpolation processing, this part high-frequency information can be recovered using neutral net.
Fig. 2 is the embodiment for obtaining the file for being used to rebuild training, and it is extracted from a collection of sample image for training
High resolution graphics area and corresponding low resolution figure area training set.Specific extracting method is as follows：Be tod using rotation, turning operation
The sample image that initial data is concentrated is extended for original 5 times, and sample image is read in into computer in the form of data matrix,
At interval of the submatrix figure area that one size of m pixel decimation is 21n*21n, wherein, m numerical value by submatrix figure area maximum
Gradient determines, so that the sampling of figure area is more sparse in low frequency part, it is more dense in HFS, specifically, to extraction
Each pixel of 21n*21n image blocks is traveled through, and calculates the four neighborhood pixels of each pixel in different directions
Difference, and be added after difference is taken absolute value, that is, calculate Obtain the gradient and frequency information of image block.The spaced pixels points of abstract image block
Mesh m takesCalculated value, so as to ensure that high frequency region obtain image block it is more dense；n
For the multiplication factor of the neutral net for training, 2 are usually chosen for.Then the subgraph area matrix of extraction is carried out n times
Down-sampling processing, obtains the low-resolution image block of 21*21 sizes corresponding to high resolution graphics area.The low resolution that will be obtained again
The bicubic interpolation that rate image block carries out n times handles to obtain the bicubic image block that size is 21n*21n.Finally by high score
Resolution figure area, low resolution figure area and bicubic figure area save as the h5 files for training.
Obtaining for rebuilding after training file, it is necessary to which image block to be inputted to the neutral net of structure.The nerve
Depth residual error network is introduced super-resolution rebuilding task, a kind of network knot for the ultra-deep level being process by network for the present invention
Structure.In the neutral net of the present invention, convolution directly is carried out to original low-resolution image, close to the position of network end-point profit
Image is up-sampled with sample level.The image that bicubic interpolation obtains and the output layer phase Calais of neutral net are predicted
Real image, substantially it is to allow the difference of neural network prediction true picture and bicubic interpolation image, that is, image
High-frequency information.Such a mode helps to reduce the training difficulty of neutral net, improves and rebuilds effect.Nerve net proposed by the present invention
Network has used 21 layers of convolutional layer and one layer of up-sampling layer, and preceding 18 layers of convolution operation is dealt with to low-resolution image,
Image resolution ratio is improved using up-sampling layer after 18 layers of convolutional layer, so can effectively improving operational speed, while upper
Three-layer coil lamination is added after sample level, ensure that the reconstruction effect of image.
Present invention introduces residual error network structure, parallel link is used between each convolutional layer, while reducing gradient disperse
Realize the fusion of multi-model.The introducing of residual error network structure lifts the depth of neutral net to 21 layers, and will study
Rate lifted to 0.1 from 10^-5, the significant increase convergence rate and precision of model.In addition, the present invention is by by interpolation image
The low-frequency information for including image maps directly to the repetition instruction that network end-point can avoid neutral net to image low-frequency information
Practice, improve the convergence rate and reconstruction precision of model.
Fig. 3 is the neural network structure in the present invention.The image of reconstruction is mainly made up of two parts, i.e.,
WhereinFor reconstruction image, X is original low-resolution image.B (X) is directly to carry out bicubic interpolation by low-resolution image to obtain
The image arrived, which represent the low frequency part of image.After C (X) is a series of convolutional layer of the low-resolution image by cascades
Output image, which represent the HFS of image.Wherein neutral net only learns to the HFS of image, drops significantly
The low training difficulty of image.
The parameter of convolutional layer represents that wherein n is the quantity of convolution kernel in convolutional layer with (n, s), and s is the size of convolution kernel,
Each convolutional layer uses 16 convolution kernels herein, and size is 3 × 3.Linear amending unit (ReLU) conduct is used after convolutional layer
Unit is activated, image passes through the output C of each convolutional layerlFor Cl=max (0, Wl*Cl-1+bl), wherein WlAnd blRespectively l
The weights of layer convolution kernel and biasing.
It is up-sampling layer behind convolutional layer, effect is to lift the resolution ratio of image.To the up-sampling layer that span is n, size
For 21*21 low-resolution image block through up-sampling layer processing after size be changed into 21n*21n.In order to reduce nerve net
The amount of calculation of network processing, up-sampling layer is disposed close to network end-point.
The optimized algorithm that the present invention uses in the training process uses the plan of variable learning rate for stochastic gradient descent method
Slightly.Initial learning rate be 0.1, often complete ten training to bulk sample sheet in training set, learning rate drop to before 10%,
This strategy can ensure suppression gradient diffusing phenomenon while model is promptly restrained, and model is converged in a preferably position
Put.
The present invention is transported after obtaining having the depth convolutional neural networks structure of residual error structure using the iteration of neutral net
Calculating realizes the continuous ratio amplification of image with pixel weighting algorithm, and solving training pattern in the prior art can not realize continuously
The problem of amplification.The present invention realizes method of the floating-point amplifier based on bicubic interpolation of image, for image is enlarged into 3.5 multiple lengths
Very little, the image after the Processing with Neural Network that low-resolution image can be passed through into twice of amplification obtains 4 again by Processing with Neural Network
The image of amplification again, then image down is target size by the kernel function of recycling bicubic interpolation.
Be illustrated in figure 4 the present invention in neutral net to image procossing the step of.Specially：, will be low in amplification process
Size gets a promotion after the depth convolutional neural networks with residual error structure that image in different resolution input training is completed, times of lifting
Number is the multiplication factor that neutral net is pre-set to input picture in training process.If after neutral net, point of image
Resolution is not reaching to target resolution, then output image is again by Processing with Neural Network.After interative computation, when output image
Resolution ratio exceedes target resolution, then carries out diminution processing to the output image, image is reached target resolution.Diminution is handled
Algorithm be similar to interpolation algorithm inverse operation, fortune is weighted to the pixel value size of image according to the ratio of image down
Calculate.
In order to verify the present invention to the validity of image enlargement processing method, on the set14 of image measurement storehouse, respectively with its
He is compared three kinds of outstanding algorithms.Fig. 5 is respectively the comparison of the image processing effect of following several algorithms：Fig. 5-1 is artwork,
Fig. 5-2 is Bicubic bicubic interpolation algorithms, and Fig. 5-3 is that A+ is improved anchor point neighbour regression algorithm, and Fig. 5-4 is for SRCNN
Super-resolution rebuilding algorithm based on convolutional neural networks, the algorithm for reconstructing of Fig. 5-5 present invention, it can be seen that in sill portion,
The visual effect of the present invention is substantially better than other method, while whole structure is relatively sharp.Fig. 6 is in image measurement storehouse set14
In, the comparison of the signal to noise ratio of each method for reconstructing under different amplification.Test result indicates that method proposed by the present invention, not only
Effect more significant than other several outstanding algorithms is all achieved in visual effect and in objective evaluation standard, is showed
Outstanding super-resolution rebuilding performance.
Finally it should be noted that：The above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although
The present invention is described in detail with reference to the foregoing embodiments, it will be understood by those within the art that：It still may be used
To be modified to the technical scheme described in previous embodiment, or equivalent substitution is carried out to which part technical characteristic；And
These modifications are replaced, and the essence of appropriate technical solution is departed from the spirit and model of various embodiments of the present invention technical scheme
Enclose.
Claims (9)
1. a kind of single image super resolution ratio reconstruction method based on depth residual error network, including：Step 1, to sample image number
Block extraction and pixel average treatment are carried out according to the image in storehouse, obtains corresponding high-resolution and low resolution training image collection
Training file；Step 2, a depth convolutional neural networks with residual error structure are built and are iterated training, and then will step
Training is iterated in the training file input depth convolutional neural networks with residual error structure obtained in rapid 1；Step
3, the data model obtained according to step 2 repetitive exercise, it is implemented in combination with by interative computation and interpolation algorithm to inputting low resolution
The continuous ratio of rate image amplifies to obtain reconstruction image.
2. the method as described in claim 1, it is characterised in that in step 1, obtain for rebuilding the training file trained
Method is, according to the submatrix figure area of the pixel gradient feature extraction corresponding size of image, to be extracted more in image HFS
Dense, low frequency part extraction is more sparse, and increases the figure area of bicubic interpolation, and the bicubic interpolation image contains image
Low frequency part.
3. method as claimed in claim 2, it is characterised in that ensure to extract more dense, low frequency part in image HFS
Extracting more sparse method is：Expand the sample image that initial data is concentrated using rotation, turning operation, by the sample image
In the form of data matrix read in computer, at interval of image block spaced pixels count out m extract a size be 21n*21n
Submatrix figure area, wherein, n be for training neutral net multiplication factor, to each picture of the 21n*21n image blocks of extraction
Vegetarian refreshments is traveled through, and calculates the difference of the four neighborhood pixels of each pixel in different directions, and difference is taken absolute value
After be added, that is, calculate The gradient and frequency information of image block are obtained, the f (x, y) can pass through 16 nearest in rectangular mesh
The weighted average of sampled point obtains, and the spaced pixels of the abstract image block m that counts out takes
Calculated value, so as to ensure that high frequency region obtain image block it is more dense.
4. the method as described in claim 1, it is characterised in that the reconstruction image is mainly made up of two parts, i.e.,WhereinFor reconstruction image, X is original low-resolution image, and B (X) is straight by low-resolution image
The image that row bicubic interpolation obtains is tapped into, which represent the low frequency part of image, C (X) is that low-resolution image passes through a system
The output image after the convolutional layer of cascade is arranged, which represent the HFS of image.
5. the method as described in claim 1, it is characterised in that in step 2, obtain the depth convolution god with residual error structure
Method through network is that depth residual error network is introduced into super-resolution rebuilding task, a kind of net for the ultra-deep level being process
Network structure.
6. method as claimed in claim 4, it is characterised in that in the depth convolutional Neural with residual error structure of the structure
In network, convolution directly is carried out to original low-resolution image, layer is up-sampled to image being utilized close to the position of network end-point
Up-sampled.
7. method as claimed in claim 4, it is characterised in that in the depth convolutional Neural with residual error structure of the structure
In network, residual error network structure is introduced, parallel link is used between each convolutional layer, multimode is realized while reducing gradient disperse
The depth of neutral net is promoted to 21 layers by the fusion of type, the introducing of residual error network structure.
8. the method as described in claim 1, it is characterised in that：In the depth convolutional Neural with residual error structure of the structure
In network, by the way that the low-frequency information for including image in interpolation image is mapped directly into network end-point, neutral net pair is avoided
The repetition training of image low-frequency information.
9. the method as described in claim 1, it is characterised in that in step 3, the continuous ratio of described pair of input low-resolution image
Example amplification procedure be：S1. size gets a promotion after low-resolution image being inputted into the neutral net, and the multiple of lifting is training
During neutral net to the multiplication factor that pre-sets of input picture；If S2. by the neutral net after, the resolution of image
Rate is not reaching to target resolution, then into step S1, if when the resolution ratio of output image exceedes target resolution, entrance step
Rapid S3；S3. diminution processing is carried out to the output image, image is reached target resolution, the algorithm of the diminution processing is similar
In the inverse operation of interpolation algorithm, the pixel value size of image is weighted according to the ratio of image down.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710426614.XA CN107358575A (en) | 2017-06-08 | 2017-06-08 | A kind of single image super resolution ratio reconstruction method based on depth residual error network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710426614.XA CN107358575A (en) | 2017-06-08 | 2017-06-08 | A kind of single image super resolution ratio reconstruction method based on depth residual error network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107358575A true CN107358575A (en) | 2017-11-17 |
Family
ID=60272460
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710426614.XA Pending CN107358575A (en) | 2017-06-08 | 2017-06-08 | A kind of single image super resolution ratio reconstruction method based on depth residual error network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107358575A (en) |
Cited By (48)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107967669A (en) * | 2017-11-24 | 2018-04-27 | 腾讯科技（深圳）有限公司 | Method, apparatus, computer equipment and the storage medium of picture processing |
| CN108175402A (en) * | 2017-12-26 | 2018-06-19 | 智慧康源（厦门）科技有限公司 | The intelligent identification Method of electrocardiogram (ECG) data based on residual error network |
| CN108235058A (en) * | 2018-01-12 | 2018-06-29 | 广州华多网络科技有限公司 | Video quality processing method, storage medium and terminal |
| CN108416736A (en) * | 2018-03-21 | 2018-08-17 | 西安邮电大学 | A kind of image super-resolution rebuilding method returned based on secondary anchor point neighborhood |
| CN108447020A (en) * | 2018-03-12 | 2018-08-24 | 南京信息工程大学 | A kind of face super-resolution reconstruction method based on profound convolutional neural networks |
| 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 |
| CN108765290A (en) * | 2018-05-29 | 2018-11-06 | 天津大学 | A kind of super resolution ratio reconstruction method based on improved dense convolutional neural networks |
| CN108765297A (en) * | 2018-06-14 | 2018-11-06 | 厦门大学 | Super resolution ratio reconstruction method based on circuit training |
| CN108765291A (en) * | 2018-05-29 | 2018-11-06 | 天津大学 | Super resolution ratio reconstruction method based on dense neural network and two-parameter loss function |
| CN108986029A (en) * | 2018-07-03 | 2018-12-11 | 南京览笛信息科技有限公司 | Character image super resolution ratio reconstruction method, system, terminal device and storage medium |
| CN109003229A (en) * | 2018-08-09 | 2018-12-14 | 成都大学 | Magnetic resonance super resolution ratio reconstruction method based on three-dimensional enhancing depth residual error network |
| CN109064408A (en) * | 2018-09-27 | 2018-12-21 | 北京飞搜科技有限公司 | A kind of method and device of multi-scale image super-resolution rebuilding |
| CN109064394A (en) * | 2018-06-11 | 2018-12-21 | 西安电子科技大学 | A kind of image super-resolution rebuilding method based on convolutional neural networks |
| CN109102463A (en) * | 2018-08-13 | 2018-12-28 | 北京飞搜科技有限公司 | A kind of super-resolution image reconstruction method and device |
| CN109345476A (en) * | 2018-09-19 | 2019-02-15 | 南昌工程学院 | High spectrum image super resolution ratio reconstruction method and device based on depth residual error network |
| CN109544451A (en) * | 2018-11-14 | 2019-03-29 | 武汉大学 | A kind of image super-resolution rebuilding method and system based on gradual iterative backprojection |
| CN109584164A (en) * | 2018-12-18 | 2019-04-05 | 华中科技大学 | Medical image super-resolution three-dimensional rebuilding method based on bidimensional image transfer learning |
| CN109685717A (en) * | 2018-12-14 | 2019-04-26 | 厦门理工学院 | Image super-resolution rebuilding method, device and electronic equipment |
| WO2019104705A1 (en) * | 2017-12-01 | 2019-06-06 | 华为技术有限公司 | Image processing method and device |
| CN109889800A (en) * | 2019-02-28 | 2019-06-14 | 深圳市商汤科技有限公司 | Image enchancing method and device, electronic equipment, storage medium |
| CN109903226A (en) * | 2019-01-30 | 2019-06-18 | 天津城建大学 | Image super-resolution rebuilding method based on symmetrical residual error convolutional neural networks |
| CN109903221A (en) * | 2018-04-04 | 2019-06-18 | 华为技术有限公司 | Image oversubscription method and device |
| CN109996023A (en) * | 2017-12-29 | 2019-07-09 | 华为技术有限公司 | Image processing method and device |
| CN110044262A (en) * | 2019-05-09 | 2019-07-23 | 哈尔滨理工大学 | Contactless precision measuring instrument and measurement method based on image super-resolution rebuilding |
| CN110060314A (en) * | 2019-04-22 | 2019-07-26 | 深圳安科高技术股份有限公司 | A kind of CT iterative approximation accelerated method and system based on artificial intelligence |
| CN110188807A (en) * | 2019-05-21 | 2019-08-30 | 重庆大学 | Tunnel pedestrian target detection method based on cascade super-resolution network and improvement Faster R-CNN |
| CN110223234A (en) * | 2019-06-12 | 2019-09-10 | 杨勇 | Depth residual error network image super resolution ratio reconstruction method based on cascade shrinkage expansion |
| CN110430419A (en) * | 2019-07-12 | 2019-11-08 | 北京大学 | A kind of multiple views naked eye three-dimensional image composition method anti-aliasing based on super-resolution |
| CN110446071A (en) * | 2019-08-13 | 2019-11-12 | 腾讯科技（深圳）有限公司 | Multi-media processing method, device, equipment and medium neural network based |
| CN110458759A (en) * | 2019-08-16 | 2019-11-15 | 杭州微算智能科技有限公司 | One kind being based on EDSR free hand drawing super resolution ratio reconstruction method |
| CN110826467A (en) * | 2019-11-22 | 2020-02-21 | 中南大学湘雅三医院 | Electron microscope image reconstruction system and method |
| CN111192215A (en) * | 2019-12-30 | 2020-05-22 | 百度时代网络技术（北京）有限公司 | Image processing method, device, equipment and readable storage medium |
| WO2020125740A1 (en) * | 2018-12-20 | 2020-06-25 | 深圳市中兴微电子技术有限公司 | Image reconstruction method and device, apparatus, and computer-readable storage medium |
| CN111583502A (en) * | 2020-05-08 | 2020-08-25 | 辽宁科技大学 | Renminbi (RMB) crown word number multi-label identification method based on deep convolutional neural network |
| CN111583143A (en) * | 2020-04-30 | 2020-08-25 | 广州大学 | Complex image deblurring method |
| CN111951164A (en) * | 2020-08-11 | 2020-11-17 | 哈尔滨理工大学 | Image super-resolution reconstruction network structure and image reconstruction effect analysis method |
| CN112183736A (en) * | 2019-07-05 | 2021-01-05 | 三星电子株式会社 | Artificial intelligence processor and method for executing neural network operation |
| CN112419146A (en) * | 2019-08-20 | 2021-02-26 | 武汉Tcl集团工业研究院有限公司 | Image processing method and device and terminal equipment |
| WO2021134874A1 (en) * | 2019-12-31 | 2021-07-08 | 深圳大学 | Training method for deep residual network for removing a moire pattern of two-dimensional code |
| CN113096011A (en) * | 2021-03-25 | 2021-07-09 | 北京达佳互联信息技术有限公司 | Image processing method and device and electronic equipment |
| CN113538235A (en) * | 2021-06-30 | 2021-10-22 | 北京百度网讯科技有限公司 | Training method and device of image processing model, electronic equipment and storage medium |
| WO2021258530A1 (en) * | 2020-06-22 | 2021-12-30 | 北京大学深圳研究生院 | Image resolution processing method, device, apparatus, and readable storage medium |
| CN114494569A (en) * | 2022-01-27 | 2022-05-13 | 光线云(杭州)科技有限公司 | Cloud rendering method and device based on lightweight neural network and residual streaming transmission |
| CN114820302A (en) * | 2022-03-22 | 2022-07-29 | 桂林理工大学 | Improved image super-resolution algorithm based on residual dense CNN and edge enhancement |
| JP2022536807A (en) * | 2019-06-18 | 2022-08-18 | ホアウェイ・テクノロジーズ・カンパニー・リミテッド | Real-time video ultra-high resolution |
| CN115190263A (en) * | 2022-09-13 | 2022-10-14 | 广州市保伦电子有限公司 | Video scaling method, device, equipment and storage medium |
| CN115409715A (en) * | 2022-11-01 | 2022-11-29 | 北京科技大学 | Hodges-Lehmann-based fire-fighting dangerous goods image super-sorting method and device |
| CN116128727A (en) * | 2023-02-02 | 2023-05-16 | 中国人民解放军国防科技大学 | Super-resolution method, system, equipment and medium for polarized radar image |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103020897A (en) * | 2012-09-28 | 2013-04-03 | 香港应用科技研究院有限公司 | Device for reconstructing based on super-resolution of multi-block single-frame image, system and method thereof |
| CN104660951A (en) * | 2015-01-21 | 2015-05-27 | 上海交通大学 | Super-resolution amplification method of ultra-high definition video image converted from high definition video image |
| CN106097253A (en) * | 2016-08-24 | 2016-11-09 | 北京印刷学院 | A kind of based on block rotation and the single image super resolution ratio reconstruction method of definition |
-
2017
- 2017-06-08 CN CN201710426614.XA patent/CN107358575A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103020897A (en) * | 2012-09-28 | 2013-04-03 | 香港应用科技研究院有限公司 | Device for reconstructing based on super-resolution of multi-block single-frame image, system and method thereof |
| CN104660951A (en) * | 2015-01-21 | 2015-05-27 | 上海交通大学 | Super-resolution amplification method of ultra-high definition video image converted from high definition video image |
| CN106097253A (en) * | 2016-08-24 | 2016-11-09 | 北京印刷学院 | A kind of based on block rotation and the single image super resolution ratio reconstruction method of definition |
Non-Patent Citations (1)
| Title |
|---|
| 孙跃文 等: ""基于深度学习的辐射图像超分辨率重建方法"", 《原子能科学技术》 * |
Cited By (75)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107967669A (en) * | 2017-11-24 | 2018-04-27 | 腾讯科技（深圳）有限公司 | Method, apparatus, computer equipment and the storage medium of picture processing |
| CN107967669B (en) * | 2017-11-24 | 2022-08-09 | 腾讯科技（深圳）有限公司 | Picture processing method and device, computer equipment and storage medium |
| US11704771B2 (en) | 2017-12-01 | 2023-07-18 | Huawei Technologies Co., Ltd. | Training super-resolution convolutional neural network model using a high-definition training image, a low-definition training image, and a mask image |
| WO2019104705A1 (en) * | 2017-12-01 | 2019-06-06 | 华为技术有限公司 | Image processing method and device |
| CN108175402A (en) * | 2017-12-26 | 2018-06-19 | 智慧康源（厦门）科技有限公司 | The intelligent identification Method of electrocardiogram (ECG) data based on residual error network |
| CN109996023A (en) * | 2017-12-29 | 2019-07-09 | 华为技术有限公司 | Image processing method and device |
| CN109996023B (en) * | 2017-12-29 | 2021-06-29 | 华为技术有限公司 | Image processing method and device |
| CN108235058A (en) * | 2018-01-12 | 2018-06-29 | 广州华多网络科技有限公司 | Video quality processing method, storage medium and terminal |
| CN108447020A (en) * | 2018-03-12 | 2018-08-24 | 南京信息工程大学 | A kind of face super-resolution reconstruction method based on profound convolutional neural networks |
| CN108416736B (en) * | 2018-03-21 | 2021-11-16 | 西安邮电大学 | Image super-resolution reconstruction method based on secondary anchor point neighborhood regression |
| CN108416736A (en) * | 2018-03-21 | 2018-08-17 | 西安邮电大学 | A kind of image super-resolution rebuilding method returned based on secondary anchor point neighborhood |
| US11593916B2 (en) | 2018-04-04 | 2023-02-28 | Huawei Technologies Co., Ltd. | Image super-resolution method and apparatus |
| WO2019192588A1 (en) * | 2018-04-04 | 2019-10-10 | 华为技术有限公司 | Image super resolution method and device |
| CN109903221B (en) * | 2018-04-04 | 2023-08-22 | 华为技术有限公司 | Image super-division method and device |
| CN109903221A (en) * | 2018-04-04 | 2019-06-18 | 华为技术有限公司 | Image oversubscription method and device |
| 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 |
| CN108765291A (en) * | 2018-05-29 | 2018-11-06 | 天津大学 | Super resolution ratio reconstruction method based on dense neural network and two-parameter loss function |
| CN108765290A (en) * | 2018-05-29 | 2018-11-06 | 天津大学 | A kind of super resolution ratio reconstruction method based on improved dense convolutional neural networks |
| CN109064394A (en) * | 2018-06-11 | 2018-12-21 | 西安电子科技大学 | A kind of image super-resolution rebuilding method based on convolutional neural networks |
| CN108765297A (en) * | 2018-06-14 | 2018-11-06 | 厦门大学 | Super resolution ratio reconstruction method based on circuit training |
| CN108765297B (en) * | 2018-06-14 | 2020-07-17 | 厦门大学 | Super-resolution reconstruction method based on cyclic training |
| CN108986029B (en) * | 2018-07-03 | 2023-09-08 | 南京览笛信息科技有限公司 | Text image super-resolution reconstruction method, system, terminal equipment and storage medium |
| CN108986029A (en) * | 2018-07-03 | 2018-12-11 | 南京览笛信息科技有限公司 | Character image super resolution ratio reconstruction method, system, terminal device and storage medium |
| CN109003229A (en) * | 2018-08-09 | 2018-12-14 | 成都大学 | Magnetic resonance super resolution ratio reconstruction method based on three-dimensional enhancing depth residual error network |
| CN109003229B (en) * | 2018-08-09 | 2022-12-13 | 成都大学 | Magnetic resonance super-resolution reconstruction method based on three-dimensional enhanced depth residual error network |
| CN109102463A (en) * | 2018-08-13 | 2018-12-28 | 北京飞搜科技有限公司 | A kind of super-resolution image reconstruction method and device |
| CN109102463B (en) * | 2018-08-13 | 2023-01-24 | 苏州飞搜科技有限公司 | Super-resolution image reconstruction method and device |
| CN109345476A (en) * | 2018-09-19 | 2019-02-15 | 南昌工程学院 | High spectrum image super resolution ratio reconstruction method and device based on depth residual error network |
| CN109064408A (en) * | 2018-09-27 | 2018-12-21 | 北京飞搜科技有限公司 | A kind of method and device of multi-scale image super-resolution rebuilding |
| CN109544451A (en) * | 2018-11-14 | 2019-03-29 | 武汉大学 | A kind of image super-resolution rebuilding method and system based on gradual iterative backprojection |
| CN109685717A (en) * | 2018-12-14 | 2019-04-26 | 厦门理工学院 | Image super-resolution rebuilding method, device and electronic equipment |
| CN109584164B (en) * | 2018-12-18 | 2023-05-26 | 华中科技大学 | Medical image super-resolution three-dimensional reconstruction method based on two-dimensional image transfer learning |
| CN109584164A (en) * | 2018-12-18 | 2019-04-05 | 华中科技大学 | Medical image super-resolution three-dimensional rebuilding method based on bidimensional image transfer learning |
| WO2020125740A1 (en) * | 2018-12-20 | 2020-06-25 | 深圳市中兴微电子技术有限公司 | Image reconstruction method and device, apparatus, and computer-readable storage medium |
| CN111353944A (en) * | 2018-12-20 | 2020-06-30 | 深圳市中兴微电子技术有限公司 | Image reconstruction method and device and computer readable storage medium |
| US20220036506A1 (en) * | 2018-12-20 | 2022-02-03 | Zte Corporation | Image reconstruction method and device, apparatus, and non-transitory computer-readable storage medium |
| US11810265B2 (en) * | 2018-12-20 | 2023-11-07 | Sanechips Technology Co., Ltd. | Image reconstruction method and device, apparatus, and non-transitory computer-readable storage medium |
| CN109903226A (en) * | 2019-01-30 | 2019-06-18 | 天津城建大学 | Image super-resolution rebuilding method based on symmetrical residual error convolutional neural networks |
| CN109903226B (en) * | 2019-01-30 | 2023-08-15 | 天津城建大学 | Image super-resolution reconstruction method based on symmetric residual convolution neural network |
| CN109889800B (en) * | 2019-02-28 | 2021-09-10 | 深圳市商汤科技有限公司 | Image enhancement method and device, electronic equipment and storage medium |
| CN109889800A (en) * | 2019-02-28 | 2019-06-14 | 深圳市商汤科技有限公司 | Image enchancing method and device, electronic equipment, storage medium |
| CN110060314A (en) * | 2019-04-22 | 2019-07-26 | 深圳安科高技术股份有限公司 | A kind of CT iterative approximation accelerated method and system based on artificial intelligence |
| CN110044262A (en) * | 2019-05-09 | 2019-07-23 | 哈尔滨理工大学 | Contactless precision measuring instrument and measurement method based on image super-resolution rebuilding |
| CN110188807A (en) * | 2019-05-21 | 2019-08-30 | 重庆大学 | Tunnel pedestrian target detection method based on cascade super-resolution network and improvement Faster R-CNN |
| CN110223234A (en) * | 2019-06-12 | 2019-09-10 | 杨勇 | Depth residual error network image super resolution ratio reconstruction method based on cascade shrinkage expansion |
| JP2022536807A (en) * | 2019-06-18 | 2022-08-18 | ホアウェイ・テクノロジーズ・カンパニー・リミテッド | Real-time video ultra-high resolution |
| JP7417640B2 (en) | 2019-06-18 | 2024-01-18 | ホアウェイ・テクノロジーズ・カンパニー・リミテッド | Real-time video ultra-high resolution |
| CN112183736A (en) * | 2019-07-05 | 2021-01-05 | 三星电子株式会社 | Artificial intelligence processor and method for executing neural network operation |
| CN110430419A (en) * | 2019-07-12 | 2019-11-08 | 北京大学 | A kind of multiple views naked eye three-dimensional image composition method anti-aliasing based on super-resolution |
| CN110430419B (en) * | 2019-07-12 | 2021-06-04 | 北京大学 | Multi-view naked eye three-dimensional image synthesis method based on super-resolution anti-aliasing |
| CN110446071A (en) * | 2019-08-13 | 2019-11-12 | 腾讯科技（深圳）有限公司 | Multi-media processing method, device, equipment and medium neural network based |
| CN110458759A (en) * | 2019-08-16 | 2019-11-15 | 杭州微算智能科技有限公司 | One kind being based on EDSR free hand drawing super resolution ratio reconstruction method |
| CN112419146A (en) * | 2019-08-20 | 2021-02-26 | 武汉Tcl集团工业研究院有限公司 | Image processing method and device and terminal equipment |
| CN112419146B (en) * | 2019-08-20 | 2023-12-29 | 武汉Tcl集团工业研究院有限公司 | Image processing method and device and terminal equipment |
| CN110826467B (en) * | 2019-11-22 | 2023-09-29 | 中南大学湘雅三医院 | Electron microscope image reconstruction system and method thereof |
| CN110826467A (en) * | 2019-11-22 | 2020-02-21 | 中南大学湘雅三医院 | Electron microscope image reconstruction system and method |
| CN111192215B (en) * | 2019-12-30 | 2023-08-29 | 百度时代网络技术（北京）有限公司 | Image processing method, device, equipment and readable storage medium |
| CN111192215A (en) * | 2019-12-30 | 2020-05-22 | 百度时代网络技术（北京）有限公司 | Image processing method, device, equipment and readable storage medium |
| WO2021134874A1 (en) * | 2019-12-31 | 2021-07-08 | 深圳大学 | Training method for deep residual network for removing a moire pattern of two-dimensional code |
| CN111583143A (en) * | 2020-04-30 | 2020-08-25 | 广州大学 | Complex image deblurring method |
| CN111583502A (en) * | 2020-05-08 | 2020-08-25 | 辽宁科技大学 | Renminbi (RMB) crown word number multi-label identification method based on deep convolutional neural network |
| WO2021258530A1 (en) * | 2020-06-22 | 2021-12-30 | 北京大学深圳研究生院 | Image resolution processing method, device, apparatus, and readable storage medium |
| CN111951164A (en) * | 2020-08-11 | 2020-11-17 | 哈尔滨理工大学 | Image super-resolution reconstruction network structure and image reconstruction effect analysis method |
| CN113096011B (en) * | 2021-03-25 | 2024-02-09 | 北京达佳互联信息技术有限公司 | Image processing method and device and electronic equipment |
| CN113096011A (en) * | 2021-03-25 | 2021-07-09 | 北京达佳互联信息技术有限公司 | Image processing method and device and electronic equipment |
| CN113538235B (en) * | 2021-06-30 | 2024-01-09 | 北京百度网讯科技有限公司 | Training method and device for image processing model, electronic equipment and storage medium |
| CN113538235A (en) * | 2021-06-30 | 2021-10-22 | 北京百度网讯科技有限公司 | Training method and device of image processing model, electronic equipment and storage medium |
| CN114494569A (en) * | 2022-01-27 | 2022-05-13 | 光线云(杭州)科技有限公司 | Cloud rendering method and device based on lightweight neural network and residual streaming transmission |
| CN114494569B (en) * | 2022-01-27 | 2023-09-19 | 光线云(杭州)科技有限公司 | Cloud rendering method and device based on lightweight neural network and residual streaming |
| CN114820302A (en) * | 2022-03-22 | 2022-07-29 | 桂林理工大学 | Improved image super-resolution algorithm based on residual dense CNN and edge enhancement |
| CN115190263A (en) * | 2022-09-13 | 2022-10-14 | 广州市保伦电子有限公司 | Video scaling method, device, equipment and storage medium |
| CN115190263B (en) * | 2022-09-13 | 2022-12-20 | 广州市保伦电子有限公司 | Video scaling method, device, equipment and storage medium |
| CN115409715A (en) * | 2022-11-01 | 2022-11-29 | 北京科技大学 | Hodges-Lehmann-based fire-fighting dangerous goods image super-sorting method and device |
| CN116128727A (en) * | 2023-02-02 | 2023-05-16 | 中国人民解放军国防科技大学 | Super-resolution method, system, equipment and medium for polarized radar image |
| CN116128727B (en) * | 2023-02-02 | 2023-06-20 | 中国人民解放军国防科技大学 | Super-resolution method, system, equipment and medium for polarized radar image |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107358575A (en) | A kind of single image super resolution ratio reconstruction method based on depth residual error network | |
| Zhu et al. | A benchmark for edge-preserving image smoothing | |
| CN104008538B (en) | Based on single image super-resolution method | |
| CN108734660A (en) | A kind of image super-resolution rebuilding method and device based on deep learning | |
| CN105976318A (en) | Image super-resolution reconstruction method | |
| CN106373109A (en) | Medical image modal synthesis method | |
| CN107527352A (en) | Remote sensing Ship Target contours segmentation and detection method based on deep learning FCN networks | |
| CN106845471A (en) | A kind of vision significance Forecasting Methodology based on generation confrontation network | |
| CN109345476A (en) | High spectrum image super resolution ratio reconstruction method and device based on depth residual error network | |
| CN110532894A (en) | Remote sensing target detection method based on boundary constraint CenterNet | |
| CN109685717A (en) | Image super-resolution rebuilding method, device and electronic equipment | |
| CN109711401A (en) | A kind of Method for text detection in natural scene image based on Faster Rcnn | |
| Chen et al. | Single image super-resolution using deep CNN with dense skip connections and inception-resnet | |
| CN107784628A (en) | A kind of super-resolution implementation method based on reconstruction optimization and deep neural network | |
| CN112801904B (en) | Hybrid degraded image enhancement method based on convolutional neural network | |
| CN108734225A (en) | A kind of transmission line construction subject image detection method based on deep learning | |
| CN111161224A (en) | Casting internal defect grading evaluation system and method based on deep learning | |
| CN112464891B (en) | Hyperspectral image classification method | |
| Kim et al. | Pynet-ca: enhanced pynet with channel attention for end-to-end mobile image signal processing | |
| CN106169174A (en) | A kind of image magnification method | |
| CN102306378A (en) | Image enhancement method | |
| Cheng et al. | DDU-Net: A dual dense U-structure network for medical image segmentation | |
| Wen et al. | A self-attention multi-scale convolutional neural network method for SAR image despeckling | |
| CN109064394B (en) | Image super-resolution reconstruction method based on convolutional neural network | |
| Ge et al. | G-Loss: A loss function with gradient information for super-resolution |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20171117 |
|
| RJ01 | Rejection of invention patent application after publication |