CN105574832A - Iteration direction filter bank based reversible depth convolution network structure - Google Patents
Iteration direction filter bank based reversible depth convolution network structure Download PDFInfo
- Publication number
- CN105574832A CN105574832A CN201510924544.1A CN201510924544A CN105574832A CN 105574832 A CN105574832 A CN 105574832A CN 201510924544 A CN201510924544 A CN 201510924544A CN 105574832 A CN105574832 A CN 105574832A
- Authority
- CN
- China
- Prior art keywords
- module
- frequency
- matrix
- network structure
- depth
- 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
- 230000002441 reversible effect Effects 0.000 title claims abstract description 22
- 238000000034 method Methods 0.000 claims abstract description 32
- 230000006798 recombination Effects 0.000 claims abstract description 23
- 238000005215 recombination Methods 0.000 claims abstract description 23
- 238000001914 filtration Methods 0.000 claims abstract description 14
- 239000011159 matrix material Substances 0.000 claims description 48
- 238000005070 sampling Methods 0.000 claims description 45
- 238000010494 dissociation reaction Methods 0.000 claims description 15
- 230000005593 dissociations Effects 0.000 claims description 15
- 238000005192 partition Methods 0.000 claims description 9
- 238000002955 isolation Methods 0.000 claims description 3
- 230000003044 adaptive effect Effects 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 3
- 238000012545 processing Methods 0.000 abstract description 3
- 238000013459 approach Methods 0.000 description 3
- 238000000354 decomposition reaction Methods 0.000 description 3
- 239000000203 mixture Substances 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000008280 blood Substances 0.000 description 1
- 210000004369 blood Anatomy 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000013332 literature search Methods 0.000 description 1
- 238000004064 recycling Methods 0.000 description 1
- 230000008521 reorganization Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000017105 transposition Effects 0.000 description 1
- 239000013598 vector Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/10—Image enhancement or restoration using non-spatial domain filtering
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
Abstract
The invention provides an iteration direction filter bank based reversible depth convolution network structure. The network structure comprises a direction filter bank module, a depth control module, a frequency recombination module and an application analysis module, wherein the direction filter bank module performs direction filtering for an input image; the depth control module controls the iteration times of the direction filter bank module; the obtained frequency resolution result is subjected to direction-and-dimension recombination in a frequency recombination module; and finally, a series of image processing related problems are solved through an analysis application module. The invention establishes the depth convolution network structure as well as provides a two-dimensional implementation method based on the iteration direction filter bank, so that the accurate reconstitution is ensured and efficiency is improved as well; an excellent processing effect for images with rich detailed information is achieved; and meanwhile, due to the frequency recombination module, the flexibility and diversification of the structure are improved, so that the network structure obtains high expansibility.
Description
Technical field
The image that the present invention relates to a kind of digital image processing field represents scheme, specifically a kind of reversible degree of depth convolutional network structure based on iteration direction bank of filters.
Background technology
Effective image representing method is most important to all kinds of image procossing application.The appearance of small echo makes the unusual performance of point of one-dimensional signal be caught in well.Then for the even higher dimensional signal of two dimension, the effect that small echo is not desirable.For 2D signal, classic method constructs separable 2-d wavelet by carrying out tensor product computing to one dimension small echo, but this is to level and vertical direction information sensing.Therefore, the signal with more complex geometry information is represented, need new method.
Through finding the literature search of prior art, a kind of method is proposed in " TheContourletTransform:AnEfficientDirectionalMultiresolu tionImageRepresentation " literary composition that MinhN.Do and MartinVetterli delivers on " IEEETransactionsonImageProcessing " (TIP) periodical of 2005, signal is divided into low frequency and high frequency two parts by laplacian pyramid by it, recycling directional filter banks carries out frequency partition to high-frequency signal, the low frequency part obtained continues to utilize laplacian pyramid and directional filter banks to repeat this process.This method shows and promotes to some extent compared with small echo in the directional information of image.Then its high-frequency information just remains unchanged after once dividing, and the meticulousr information made like this can not be represented well." InvariantScatteringConvolutionNetworks " that JoanBruna and StephaneMallat delivers on " IEEETransactionsonPatternAnalysisandMachineIntelligence " periodical in 2013 one proposes a kind of both scatternets for classification and identification problem in literary composition.This network is become with average operation group by a series of wavelet filter, modulo operation, and every one deck exports a low frequency subgraph and multiple high frequency subgraph, and each high frequency subgraph continues to decompose at lower one deck.But this network structure is after ground floor obtains low frequency subgraph, just no longer decomposes further it, and this network is discussed based on continuous domain, very not desirable discretization method, and does not have Accurate Reconstruction.
Summary of the invention
The present invention is directed to the deficiencies in the prior art, provide a kind of based on iteration direction bank of filters reversible degree of depth convolutional network structure, can low frequency and radio-frequency component repeatedly be divided and be recombinated, and the Digital Implementation method provided easily and effectively, also can be applied to multiple field further as a kind of general image conversion method.
The present invention is achieved by the following technical solutions:
The invention provides a kind of reversible degree of depth convolutional network structure based on iteration direction bank of filters, comprise: directional filter banks module, depth control block, frequency recombination module and analytical applications module, described directional filter banks module and described depth control block determine the isolation of input picture jointly, wherein:
Described directional filter banks module utilizes sampling matrix and directional filter banks to the filtering of input picture travel direction, and sampling matrix and anisotropic filter alternate cycles obtain coefficient of dissociation and export to described frequency recombination module;
The cycle index of described depth control block to direction filter bank block controls;
Described frequency recombination module carries out frequency based on the coefficient of dissociation that described directional filter banks module exports and reconfigures, and the new frequency partition result after restructuring is outputted to described analytical applications module;
The coefficient of dissociation with new frequency distribution that described analytical applications module receive frequency recombination module exports, and be further processed to solve application problem to this coefficient.
Preferably, described directional filter banks module and described depth control block determine the isolation of input picture jointly, wherein: described directional filter banks module comprises a fan-shaped directional filter banks and two sampling matrixs, they repeatedly decompose input picture iteratively, namely obtain the repeatedly division of the frequency domain of input picture; Decompose and specifically comprise the steps:
Step one, is input to the fan-shaped directional filter banks of binary channels and sampling matrix Q by image
0, obtain two groups of matrix of coefficients, tackle the frequency partition of horizontal and vertical directions respectively;
Step 2, by the output of the step one input fan-shaped directional filter banks of binary channels and sampling matrix Q
1, carry out filtering and down-sampling process, obtain four groups of matrix of coefficients, respectively corresponding four direction frequency partition;
Step 3, repetition step one and step 2, until reach the stop condition of depth control block setting, obtain whole coefficient of dissociation.
More preferably, described sampling matrix and fan-shaped directional filter banks alternate cycles number of times are determined by the restrictive condition of depth control block, but need ensure degree of depth l >=2.
More preferably, described sampling matrix adopts non-diagonal sampling matrix, and fan-shaped directional filter banks, utilizes McClellan transform that one dimension biorthogonal two-channel PR filter banks is mapped as inseparable two-dimentional biorthogonal compactly supported wavelets; After completing steps three, each quadrant of corresponding frequency domain has similar division to the frequency domain obtained after step one.
Preferably, described depth control block, its convolutional network degree of depth controlled is given by pre-setting preset parameter, also can provide Rule of judgment by connected applications, form adaptive degree of depth convolutional network, need ensure degree of depth l >=2.
Preferably, described directional filter banks module, i-th (i≤l) layer decomposes the frequency domain obtained and is divided in each 1/2
i/2the frequency domain that sub-block and the second layer obtain divides has similar form; And obtain all 2
iindividual area block has same shape, is all isosceles right triangle.These two kinds of similaritys ensure that the simple high efficiency of total.
Preferably, described frequency recombination module, designs different recombination forms by the statistics of each subgraph of input being carried out to significant coefficient according to the object of analytical applications module, obtains new frequency domain and divide.
More preferably, described frequency recombination module, is reconfigured by the non-zero number and significance level calculating the coefficient of dissociation under the different directions different scale that finally obtains, to reduce computation complexity and optimum results.
The reversible degree of depth convolutional network structure based on iteration direction bank of filters adopted in the present invention is that image procossing provides a kind of new method for expressing.Realize the discrete of network by the combination of non-diagonal sampling matrix and inseparable two-dimensional directional bank of filters, this has the accurate separability of better frequency domain and perfect reconstruction than directly carrying out rotation process to wave filter.The similarity utilizing each sub-block frequency domain to divide completes meticulousr frequency domain division by the iteration of two groups of sampling matrixs and bank of filters.The statistics of the coefficient of dissociation utilizing iterative filter group to export carries out frequency restructuring, is adapted to different applications better to make it.
Compared with prior art, the present invention has following beneficial effect:
The present invention establishes new degree of depth convolutional network structure, and described directional filter banks module adopts the two-dimensional discrete implementation method based on iteration direction bank of filters, ensure that and calculates simplicity and Accurate Reconstruction.The present invention utilizes the frequency distribution similarity of each sub-block of frequency domain repeatedly to divide low frequency and radio-frequency component, improves the capturing ability to enriching detailed information.By pull-in frequency reorganization operation, improve dirigibility and the extensibility of total.
Accompanying drawing explanation
By reading the detailed description done non-limiting example with reference to the following drawings, other features, objects and advantages of the present invention will become more obvious:
Fig. 1 is the structured flowchart of present system one embodiment;
Fig. 2 is the fundamental diagram of iteration direction filter bank block and depth control block.
Embodiment
Below in conjunction with specific embodiment, the present invention is described in detail.Following examples will contribute to those skilled in the art and understand the present invention further, but not limit the present invention in any form.It should be pointed out that to those skilled in the art, without departing from the inventive concept of the premise, some distortion and improvement can also be made.These all belong to protection scope of the present invention.
As shown in Figure 1, a kind of reversible degree of depth convolutional network structure based on iteration direction bank of filters, comprise: directional filter banks module and depth control block, frequency recombination module and analytical applications module, wherein: directional filter banks module utilizes fan-shaped directional filter banks to the filtering of input picture travel direction, the cycle index of depth control block to direction filter bank block controls, the frequency resolution result obtained travel direction and yardstick in frequency recombination module reconfigure, and finally solve a series of images process relevant issues for analytical applications module.
As shown in Figure 2, described directional filter banks module comprises a fan-shaped directional filter banks and two sampling matrixs, and they repeatedly decompose input picture iteratively, namely obtains the repeatedly division of the frequency domain of input picture; Described directional filter banks module and depth control block realize this degree of depth convolutional network jointly, and two groups of sampling matrixs and fan-shaped directional filter banks alternate cycles divide frequency domain, along with the increase of the degree of depth, obtain meticulousr yardstick and the decomposition result in direction.
Described directional filter banks module, decomposes and specifically comprises the steps:
Step one, is input to the fan-shaped directional filter banks of binary channels and sampling matrix Q by image
0, obtain two groups of matrix of coefficients, tackle the frequency partition of horizontal and vertical directions respectively;
Step 2, by the output of the step one input fan-shaped directional filter banks of binary channels and sampling matrix Q
1, carry out filtering and down-sampling process, obtain four groups of matrix of coefficients, respectively corresponding four direction frequency partition;
Step 3, repetition step one and step 2, until reach the stop condition of depth control block setting, obtain whole coefficient of dissociation.
In the present embodiment, described sampling matrix is respectively Q
0, Q
1, the frequency domain representation of fan-shaped wave-wave device is H (ω), wherein ω=(ω
1, ω
1)
tfor two-dimentional angular frequency; To original input image x, through first group of binary channels fan-filter H
i(ω) (i gets 0,0 passage of the corresponding two channels filter of 1 difference and 1 passage) and sampling matrix Q
0after the frequency domain representation that obtains be:
Wherein: N (Q
0 t) for there is shape as Q
0 tt, t ∈ (0,1)
2the set of integer vectors, Q
0 -Tfor sampling matrix Q
0the transposition of inverse matrix, m gets set N (Q
0 t) in element; Repeat the decomposition result that this step obtains more deep layer.
The hyperchannel of a l level network is numbered:
Wherein, t
tfor the channel number of the fan-shaped directional filter banks that the t layer that this passage is corresponding passes through, t
rget 0 or 1.
In the present embodiment, according to the exchange regulation of sampling with filtering, concrete:
Use sampling matrix Q
ithrough binary channels fan-filter H after (i gets 0,1) down-sampling
i(ω) filtering, is equivalent to and first uses filters H
i(M
tω) sampling matrix Q is used in filtering again
idown-sampling, H
i(Q
i tω) pass through binary channels fan-filter H
i(ω) with sampling matrix Q
icarry out up-sampling to obtain.Exchange the order of sampling matrix and wave filter successively, obtain the sampling matrix of the final equivalence of passage k
with the wave filter of equivalence
for:
The decomposition result of the passage k finally obtained
frequency domain representation be:
In the present embodiment, the stop condition of depth control block is set as: l=6, then the step one in directional filter banks module and step 2 repeat 3 times respectively, export 64 coefficient of dissociation matrixes
frequency domain is divided into 64 pieces.
In the present embodiment, frequency recombination module is to all coefficient of dissociation
value add up, by arranging from big to small, the value of M (0<M<N, N are image slices vegetarian refreshments number) individual larger coefficient before retaining, the value of all the other coefficients is set to 0.After carrying out this step process, the size constancy of each matrix of coefficients, only has partial value to become 0, obtains the new coefficient that frequency recombination module exports
In the present embodiment, analytical applications module is to matrix of coefficients
rebuild, obtain the approximate image of initial input picture.Concrete implementation step is:
Step one, to matrix of coefficients
with sampling matrix Q
i(i=lmod2+1) up-sampling is carried out, then with wave filter G
j(ω) (j=kmod2) carries out filtering, realizes the interpolation operation to matrix of coefficients.Its median filter G
j(ω) be decomposable process filters H used
j(ω) spatial domain reversion, i.e. G
j(ω)=H
j(-ω).
Step 2, the adjacent coefficient matrix after step one being processed from passage 0 is added between two, obtains 2
l-1individual matrix of coefficients
Step 3, upgrades l value: l=l-1, if l ≠ 0 after upgrading, then repeats step one and step 2; If l=0, obtain net result
for the approximate image of original input image x.
Implementation result
Analytical applications module in the present embodiment is set to None-linear approximation, and its implementation procedure adopts the step in foregoing invention content to implement.Employing is of a size of gray scale picture " barabara.png " (512 × 512) and tests.Relatively adopt the methods of people in " TheContourletTransform:AnEfficientDirectionalMultiresolu tionImageRepresentation " such as the reversible degree of depth convolutional network structural approach based on iteration direction bank of filters of the present invention and MinhN.Do, and traditional 2-d discrete wavelet method.Three kinds of methods all adopt 6553 most important coefficients to approach.
Its result is: adopt above-mentioned three kinds of methods respectively, and the Y-PSNR of what the present embodiment obtained approach image is 26.92 decibels, and profile ripple and wavelet method obtain result and be respectively 26.46 decibels and 26.03 decibels.Experiment shows, the reversible degree of depth convolutional network structure based on iteration direction bank of filters that the present embodiment proposes has better effect.
Above specific embodiments of the invention are described.It is to be appreciated that the present invention is not limited to above-mentioned particular implementation, those skilled in the art can make various distortion or amendment within the scope of the claims, and this does not affect flesh and blood of the present invention.
Claims (10)
1. the reversible degree of depth convolutional network structure based on iteration direction bank of filters, it is characterized in that, comprise: directional filter banks module, depth control block, frequency recombination module and analytical applications module, described directional filter banks module and described depth control block determine the isolation of input picture jointly, wherein:
Described directional filter banks module utilizes sampling matrix and directional filter banks to the filtering of input picture travel direction, and sampling matrix and anisotropic filter alternate cycles obtain coefficient of dissociation and export to described frequency recombination module;
The cycle index of described depth control block to direction filter bank block controls;
Described frequency recombination module carries out frequency based on the coefficient of dissociation that described directional filter banks module exports and reconfigures, and the new frequency partition result after restructuring is outputted to described analytical applications module;
The coefficient of dissociation with new frequency distribution that described analytical applications module receive frequency recombination module exports, and be further processed to solve application problem to this coefficient.
2. the reversible degree of depth convolutional network structure based on iteration direction bank of filters according to claim 1, it is characterized in that, described directional filter banks module comprises a fan-shaped directional filter banks and two sampling matrixs, they repeatedly decompose input picture iteratively, namely obtain the repeatedly division of the frequency domain of input picture; Described decomposable process specifically comprises the steps:
Step one, is input to the fan-shaped directional filter banks of binary channels and sampling matrix Q by image
0, carry out filtering and down-sampling process, obtain two groups of matrix of coefficients, tackle the frequency partition of horizontal and vertical directions respectively;
Step 2, is input to the fan-shaped directional filter banks of binary channels and sampling matrix Q by the output of step one
1, carry out filtering and down-sampling process, obtain four groups of matrix of coefficients, respectively corresponding four direction frequency partition;
Step 3, repeats step one and step 2 until reach the stop condition of depth control block setting to each group matrix of coefficients, obtains whole coefficient of dissociation.
3. the reversible degree of depth convolutional network structure based on iteration direction bank of filters according to claim 2, it is characterized in that, the number of times of described sampling matrix and fan-shaped directional filter banks alternate cycles is determined by the restrictive condition of depth control block, and degree of depth l >=2.
4. the reversible degree of depth convolutional network structure based on iteration direction bank of filters according to claim 2, is characterized in that, described sampling matrix adopts non-diagonal sampling matrix; Described fan-shaped directional filter banks, utilizes McClellan transform that one dimension biorthogonal two-channel PR filter banks is mapped as inseparable two-dimentional biorthogonal compactly supported wavelets; After completing steps three, each quadrant of corresponding frequency domain has similar division to the frequency domain obtained after step one.
5. the reversible degree of depth convolutional network structure based on iteration direction bank of filters according to claim 1, it is characterized in that, described depth control block controls the convolutional network degree of depth, this degree of depth provides Rule of judgment by pre-setting the given or connected applications of preset parameter, form adaptive degree of depth convolutional network, and ensure degree of depth l >=2.
6. the reversible degree of depth convolutional network structure based on iteration direction bank of filters according to claim 1, it is characterized in that, described frequency recombination module, by adding up coefficient of dissociation, and design different recombination forms according to the object of analytical applications module, obtain new frequency domain and divide.
7. the reversible degree of depth convolutional network structure based on iteration direction bank of filters according to claim 6, it is characterized in that, the value of described frequency recombination module to all coefficient of dissociation is added up, by arranging from big to small, the value of M larger coefficient before retaining, the value of all the other coefficients is set to 0, after carrying out this step process, the size constancy of each matrix of coefficients, only has partial value to become 0.
8. the reversible degree of depth convolutional network structure based on iteration direction bank of filters according to any one of claim 1-7, it is characterized in that, described decomposable process is reversible, its inverse system is operated by the up-sampling of iteration and filtering operation realizes, wherein sampling matrix is identical with decomposable process sampling matrix, and wave filter used herein and decomposable process directional filter banks median filter used reverse in spatial domain each other.
9. the reversible degree of depth convolutional network structure based on iteration direction bank of filters according to any one of claim 1-7, is characterized in that, described directional filter banks module, i-th (i≤l) layer decomposes the frequency domain obtained and is divided in each 1/2
i/2the frequency domain that sub-block and the second layer obtain divides has similar form; And obtain all 2
iindividual area block has same shape, is all isosceles right triangle.
10. the reversible degree of depth convolutional network structure based on iteration direction bank of filters according to any one of claim 1-7, it is characterized in that, described analytical applications module is rebuild frequency recombination module output coefficient, obtain initial input picture approximate image, concrete implementation step is:
Step one, to matrix of coefficients
with sampling matrix Q
icarry out up-sampling, i=lmod2+1, mod are remainder operation, then with wave filter G
j(ω) carry out filtering, j=kmod2, realize the interpolation operation to matrix of coefficients, its median filter G
j(ω) be decomposable process filters H used
j(ω) spatial domain reversion, i.e. G
j(ω)=H
j(-ω);
Step 2, the adjacent coefficient matrix after step one being processed from passage 0 is added between two, obtains 2
l-1individual matrix of coefficients
k=0,1 ..., 2
l-1-1;
Step 3, upgrades l value: l=l-1, if l ≠ 0 after upgrading, then repeats step one and step 2; If l=0, then obtain net result
for the approximate image of original input image x.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510924544.1A CN105574832B (en) | 2015-12-11 | 2015-12-11 | Based on the reversible depth convolutional network structure of iteration direction filter group |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510924544.1A CN105574832B (en) | 2015-12-11 | 2015-12-11 | Based on the reversible depth convolutional network structure of iteration direction filter group |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105574832A true CN105574832A (en) | 2016-05-11 |
| CN105574832B CN105574832B (en) | 2019-11-15 |
Family
ID=55884927
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510924544.1A Active CN105574832B (en) | 2015-12-11 | 2015-12-11 | Based on the reversible depth convolutional network structure of iteration direction filter group |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105574832B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108416805A (en) * | 2018-03-12 | 2018-08-17 | 中山大学 | A kind of intrinsic image decomposition method and device based on deep learning |
| CN108805030A (en) * | 2018-05-15 | 2018-11-13 | 成都理想境界科技有限公司 | A kind of feature extracting method and system applied to deep learning |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102142136A (en) * | 2011-03-05 | 2011-08-03 | 河海大学常州校区 | Neural network based sonar image super-resolution reconstruction method |
| CN102163329A (en) * | 2011-03-15 | 2011-08-24 | 河海大学常州校区 | Super-resolution reconstruction method of single-width infrared image based on scale analogy |
| CN101566688B (en) * | 2009-06-05 | 2012-02-08 | 西安电子科技大学 | Method for reducing speckle noises of SAR image based on neighborhood directivity information |
| CN102389309A (en) * | 2011-07-08 | 2012-03-28 | 首都医科大学 | Compressed sensing theory-based reconstruction method of magnetic resonance image |
| CN102879822A (en) * | 2012-09-28 | 2013-01-16 | 电子科技大学 | Contourlet transformation based seismic multi-attribute fusion method |
| CN102298768B (en) * | 2010-06-24 | 2013-10-02 | 江南大学 | High-resolution image reconstruction method based on sparse samples |
| CN103345732A (en) * | 2013-07-26 | 2013-10-09 | 电子科技大学 | Pulse coupled neural network (PCNN) image enhancement algorithm and device based on Contourlet transformation |
| CN104766290A (en) * | 2015-03-27 | 2015-07-08 | 哈尔滨工程大学 | Pixel information estimation fusion method based on quick NSCT |
-
2015
- 2015-12-11 CN CN201510924544.1A patent/CN105574832B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101566688B (en) * | 2009-06-05 | 2012-02-08 | 西安电子科技大学 | Method for reducing speckle noises of SAR image based on neighborhood directivity information |
| CN102298768B (en) * | 2010-06-24 | 2013-10-02 | 江南大学 | High-resolution image reconstruction method based on sparse samples |
| CN102142136A (en) * | 2011-03-05 | 2011-08-03 | 河海大学常州校区 | Neural network based sonar image super-resolution reconstruction method |
| CN102163329A (en) * | 2011-03-15 | 2011-08-24 | 河海大学常州校区 | Super-resolution reconstruction method of single-width infrared image based on scale analogy |
| CN102389309A (en) * | 2011-07-08 | 2012-03-28 | 首都医科大学 | Compressed sensing theory-based reconstruction method of magnetic resonance image |
| CN102879822A (en) * | 2012-09-28 | 2013-01-16 | 电子科技大学 | Contourlet transformation based seismic multi-attribute fusion method |
| CN103345732A (en) * | 2013-07-26 | 2013-10-09 | 电子科技大学 | Pulse coupled neural network (PCNN) image enhancement algorithm and device based on Contourlet transformation |
| CN104766290A (en) * | 2015-03-27 | 2015-07-08 | 哈尔滨工程大学 | Pixel information estimation fusion method based on quick NSCT |
Non-Patent Citations (5)
| Title |
|---|
| JOAN BRUNA 等: "Invariant Scattering Convolution Networks", 《PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
| MINH N. DO，MARTIN VETTERLI: "CONTOURLETS: A DIRECTIONAL MULTIRESOLUTION IMAGE REPRESENTATION", 《SIGNALS, SYSTEMS AND COMPUTERS》 * |
| 冯燕: "高光谱图像压缩技术研究", 《中国博士学位论文全文数据库 信息科技辑》 * |
| 刘晓山: "光照变化条件下人脸识别技术研究", 《中国博士学位论文全文数据库 信息科技辑》 * |
| 叶传奇 等: "基于区域分割的NSCT域SAR与全色图像融合算法", 《***工程与电子技术》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108416805A (en) * | 2018-03-12 | 2018-08-17 | 中山大学 | A kind of intrinsic image decomposition method and device based on deep learning |
| CN108416805B (en) * | 2018-03-12 | 2021-09-24 | 中山大学 | Intrinsic image decomposition method and device based on deep learning |
| CN108805030A (en) * | 2018-05-15 | 2018-11-13 | 成都理想境界科技有限公司 | A kind of feature extracting method and system applied to deep learning |
| CN108805030B (en) * | 2018-05-15 | 2022-03-08 | 成都理想境界科技有限公司 | Feature extraction method and system applied to deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105574832B (en) | 2019-11-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110660038B (en) | Multispectral image and full-color image fusion method based on generation countermeasure network | |
| Cui et al. | Image compressed sensing using non-local neural network | |
| CN106373167A (en) | Compressed sensing nuclear magnetic resonance imaging method based on deep neural network | |
| CN104574336A (en) | Super-resolution image reconstruction system based on self-adaptation submodel dictionary choice | |
| CN110533591B (en) | Super-resolution image reconstruction method based on codec structure | |
| CN104285239A (en) | Image processing device, image processing method, program, print medium, and recording medium | |
| CN101847256A (en) | Image denoising method based on adaptive shear wave | |
| Kavitha et al. | Medical image fusion using combined discrete wavelet and ripplet transforms | |
| Nguyen-Duc et al. | Frequency-splitting dynamic MRI reconstruction using multi-scale 3D convolutional sparse coding and automatic parameter selection | |
| CN104700440B (en) | Magnetic resonant part K spatial image reconstruction method | |
| CN103500444A (en) | Polarization image fusion method | |
| CN109146792A (en) | Chip image super resolution ratio reconstruction method based on deep learning | |
| Datta et al. | Efficient interpolated compressed sensing reconstruction scheme for 3D MRI | |
| CN105976334A (en) | Three-dimensional filtering denoising algorithm based denoising processing system and method | |
| CN105574832A (en) | Iteration direction filter bank based reversible depth convolution network structure | |
| CN102855644B (en) | Large-scale topographic data compression and decompression method based on wavelet self-feedback | |
| Mahmoud | A smart single matrix realization of fast walidlet transform | |
| CN116310452B (en) | Multi-view clustering method and system | |
| Zhao et al. | Adaptive intelligent single particle optimizer based image de-noising in shearlet domain | |
| CN103236041A (en) | Image super resolution reconstruction method on basis of Contourlet transformation | |
| Xiao et al. | MFEN: Lightweight multi-scale feature extraction super-resolution network in embedded system | |
| Haq et al. | Block-based compressed sensing of MR images using multi-rate deep learning approach | |
| CN114529482A (en) | Image compressed sensing reconstruction method based on wavelet multi-channel depth network | |
| CN103747268B (en) | A kind of layered self-adapting threshold video denoising method | |
| Shoberg et al. | The multiresolution analysis with radial symmetry elements |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |