CN109785249A - A kind of Efficient image denoising method based on duration memory intensive network - Google Patents
A kind of Efficient image denoising method based on duration memory intensive network Download PDFInfo
- Publication number
- CN109785249A CN109785249A CN201811576192.5A CN201811576192A CN109785249A CN 109785249 A CN109785249 A CN 109785249A CN 201811576192 A CN201811576192 A CN 201811576192A CN 109785249 A CN109785249 A CN 109785249A
- Authority
- CN
- China
- Prior art keywords
- network
- noise
- rdu
- block
- module
- 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
- 238000000034 method Methods 0.000 title claims abstract description 52
- 238000000605 extraction Methods 0.000 claims abstract description 44
- 238000012549 training Methods 0.000 claims abstract description 39
- 230000007850 degeneration Effects 0.000 claims abstract description 24
- 238000012360 testing method Methods 0.000 claims abstract description 12
- 238000011156 evaluation Methods 0.000 claims abstract description 11
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims abstract description 8
- 230000008569 process Effects 0.000 claims description 19
- 238000001228 spectrum Methods 0.000 claims description 15
- 239000000284 extract Substances 0.000 claims description 13
- 238000013507 mapping Methods 0.000 claims description 7
- 230000002708 enhancing effect Effects 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 230000010354 integration Effects 0.000 claims description 3
- 239000000463 material Substances 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims description 3
- 230000002085 persistent effect Effects 0.000 claims description 3
- 238000011946 reduction process Methods 0.000 claims description 3
- 241000208340 Araliaceae Species 0.000 claims description 2
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 claims description 2
- 235000003140 Panax quinquefolius Nutrition 0.000 claims description 2
- 235000008434 ginseng Nutrition 0.000 claims description 2
- 239000012141 concentrate Substances 0.000 claims 1
- 238000012545 processing Methods 0.000 abstract 1
- 230000000694 effects Effects 0.000 description 7
- 238000013135 deep learning Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000010207 Bayesian analysis Methods 0.000 description 1
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The present invention relates to the Efficient image denoising methods based on duration memory intensive network, belong to field of image processing.The present invention chooses training set, verifying collection and test set first, and is pre-processed, and the degeneration block with noise and the clear block without noise are obtained；The degeneration block with noise is handled by shallow-layer feature extraction module, duration memory intensive network module, internal layer nested networks module and residual error evaluation module, obtains estimating denoising block；Denoising block and the known clear block calculating mean square error without noise are estimated using what network exported, to carry out loss value metric；Since network end-point, weight gradient value is sought using Adam operator and updates network parameter；Training terminates and obtains denoising model after network convergence；The image of removal noise is obtained containing gaussian noise image in the input of trained denoising model.The present invention can handle white Gaussian noise present in natural image in the case where not destroying original texture image marginal information with degree of precision.
Description
Technical field
The present invention relates to a kind of Efficient image denoising methods based on duration memory intensive network, belong to image procossing skill
Art field.
Background technique
Image denoising is a classical basic problem in computer vision, and the problem is in academic research and industrial application
In obtain extensive concern.The target of denoising is that clear image x is recovered from a noisy image y=x+n, therein to make an uproar
Sound n is often assumed to be additive white Gaussian noise.From the point of view of Bayesian analysis, under the conditions of known to the likelihood value, image
Prior model plays quite crucial effect to denoising.In past many decades, denoising method has enumerated most figure
As priori, including non local self-similarity model, sparse model, gradient former and Hidden Markov domain model etc., wherein being based on
The method of self similarity model, such as the denoising effect of BM3D show well.
However, existing for the denoising model based on transcendental method one be obviously disadvantageous in that, test phase need compared with
The long time optimizes calculating, and limited in view of the breakthrough of original technology basic theory, and the denoising performance of image arrived bottle
In the neck stage, further to increase substantially recovery effect just becomes to be increasingly difficult to.Simultaneously as denoising model is non-convex problem,
The some parameters being related in optimization process are manually set, this also further constrains further mentioning for denoising performance
It is high.In recent years, occurred in succession in engineering fields such as image classification, identifications compared with quantum jump, academia with depth learning technology
Just attempt that depth convolutional network is used in combination to carry out image denoising, and introduces the calculation outstanding of some performances in heuristic process
Method.
2008, Viren Jain et al. proposed to use method (the Natural Image Denoising of convolutional network
With Convolutional Networks, abbreviation CN1) Lai Jinhang natural image denoising, this method combine specific noise
Model generates training sample, by training, obtain better than based on wavelet method to based on the related of markov random file
Method, the performance in blind denoising are equally matched with other models based on non-blind setting.Author thinks, based on convolutional network with
Method based on markov random file (MRF) has certain connection on mathematical notation, but MRF method is in the ginseng of probabilistic model
A large amount of computing resource is occupied in number estimation and deduction.The method of convolutional network then avoids carrying out the statistical of density estimation
Analysis, but regression problem is converted by denoising process.2012, Xie Junyuan et al. proposed to stack sparse noise certainly
Coding method (Stacked sparse denoising auto-encoder architecture, abbreviation SSDA), this method
Sparse coding and depth network is used in combination to carry out image denoising, wherein the denoising of depth Web vector graphic pre-training is next from encoding
It is initialized.Similar to the training process of CN1, SSDA successively increases network concealed layer, and is joined using the weight of front multilayer
It counts to initialize the initial parameter value of next hidden layer.And after one hidden layer of every increase, it introduces KL divergence and works as to calculate
The penalty values of preceding all layer state drags make the mean value activation value of hidden unit most by using lesser logarithm weight
It is sparse to achieve the effect that may tend to zero.The network of both the above method is successively trained, and at the beginning of the weight of network
Initial value not automatically generates, but requires human intervention every layer of inside, which can constrain master mould to a certain extent
Performance is optimal value, and whole training process is time-consuming.Meanwhile the construction of model is all layer-by-layer in fully-connected network
It stacks, this will promote the population parameter for generating model also abnormal big when network depth is shallower.
2016, Mao Xiaojiao et al. proposed that the depth convolutional encoding decoding network RED30 of symmetrical parallel link is used
In image denoising, the network is by multiple convolutional layers and the layer that deconvolutes as basic unit, and directly progress is from noise image to clear
End-to-end study between image.Due to having used symmetrical connection, picture signal is directly passed in the process of backpropagation
It pulls over the closer shallow-layer of distance input image, therefore can largely accelerate the training of network entirety, and alleviate ladder well
Spend disappearance problem；Simultaneously because convolution process is directly to carry out information extraction from original image, and the layer that deconvolutes is responsible for extraction
Feature is rebuild, therefore the direct energy of flow between information avoids the loss of original image signal to the full extent.It is opened from the network
Begin, neural network relevant to image denoising starts gradually to develop to deeper.2017, Zhang Kai et al. was proposed
DnCNN, this method are normalized to accelerate network training using residual error study and batch.Wherein the application of residual error learning strategy makes net
Network only learns noise information, obtains final output by the difference between input value and noise information.Above two algorithm exists
Excellent properties are shown on image denoising, but do not make full use of the characteristic information of image in finite layer, and are only to make
Simple planar, stacked is carried out with convolutional layer or is stacked using residual block.The chain type connection structure of basic block is to certain depth
After the problem of being easy to cause gradient to disappear, and its depth deficiency is difficult to obtain a complicated mapping process.For this purpose, of the invention
Existing mutual restricting relation between the two factors is released by introducing dense network structure.
Summary of the invention
The present invention provides a kind of Efficient image denoising methods based on duration memory intensive network, are able to achieve image
Efficiently denoising, while retaining the details of original image as much as possible.
The technical scheme is that a kind of Efficient image denoising method based on duration memory intensive network
(Hierarchical Nesting Dense Block for Single-image Super-resolution, referred to as
DBSR), include the following steps:
Step1, training set, verifying collection and test set are chosen, and training set and verifying collection is pre-processed, obtained band and make an uproar
The degeneration block of sound and the clear block without noise；
Step2, by shallow-layer feature extraction module, duration memory intensive network module, internal layer nested networks module and
Residual error evaluation module handles the degeneration block with noise, obtains estimating denoising block；
Preliminary noise feature extraction is carried out to input picture using shallow-layer feature extraction module first, will then extract gained
Information be transmitted to duration memory intensive network module, carry out fine noise using the stacking of residual error dense cell by the module
The extraction of feature；After the three-dimensional arrangement of multiple residual error dense cells output, gained information is further transferred to two accesses
Internal layer nested networks module carries out feature and refines process；Obtained information is the making an uproar of obtaining of extraction for being network at this time
Sound subtract each other can be obtained and estimates denoising by the output of shallow-layer feature extraction module and the output valve of internal layer nested networks module
Block；
Step3, using network export estimate denoising block and the known clear block without noise calculates mean square error, with into
Row loss value metric；
Step4, since network end-point, seek weight gradient value using Adam operator and update network parameter；
Step5, circulation step Step2-Step4, training terminates and obtains denoising model after network convergence；
Step6, the image of removal noise is obtained containing gaussian noise image in the input of trained denoising model.
Further, in the step Step1,900 2K used in image restoration and enhancing match NTIRE are selected
Training set, verifying collection of the high definition png image as model, wherein preceding 800 images are used for model training, latter 100 for testing
Card；It selects image denoising standard data set Set12, BSD68 to test network model, is concentrated in training set and verifying, it is right
Every 2K high definition original image adds white Gaussian noise, is then intercepted respectively in high definition original image and the same position of noise pattern with step-length 27
The image block that size is 96 × 96, obtains the clear block y without noisenfWith the degeneration block x with noise.
Further, in the step 2, the network first floor is preliminary using the shallow-layer feature extraction module comprising single convolution block
Extract noise characteristic；Then, it using residual error dense cell as basic unit, and is held by the layer-by-layer nesting of basic unit to construct
Continuous property memory intensive network module, to extract more granularity noise characteristic information；In the latter half of network, internal layer nested networks make
Channel integration is carried out to noise characteristic map with two parallel sub-networks, is realized by different numbers, different size of convolution kernel
The fining of noise information feature, to obtain image high-frequency noise information；Finally, network end-point will include high-frequency noise information
Picture material is integrated with the noise characteristic tentatively extracted, by asking the way of difference to remove noise in residual error evaluation module
Image.
Further, the specific steps of the step Step2:
Step2.1, shallow-layer feature extraction module tentatively extract noise characteristic: using the degeneration block x with noise as input value,
Increase the port number of input block by shallow-layer feature extraction module, and carries out the preliminary extraction of noise information to it；
Shallow-layer feature extraction module using the degeneration block x with noise as input, by convolution operation Conv, weight bias and
After band parameters revision linear unit PReLU, T is exported-1；The process is expressed with mathematical formulae are as follows: T-1=PReLU (aC1,C1),
Middle C1=W1*x+b1The output of process, W are biased for convolution1For the weight parameter of convolution operation, b1For amount of bias, a is that band parameter is repaired
The initial value of linear positive unit PReLU；
Step2.2, duration memory intensive network module refine noise characteristic information:
Duration memory intensive network module (persistent memory dense block, PMDB) is intensive with residual error
Unit (Residual Dense Unit, RDU) is used as its basic unit, passes through the way of nesting RDU further inside RDU
To achieve the purpose that extract more granularity noise characteristic information；
Specific practice are as follows: PMDB is with the output valve T of shallow-layer feature extraction module-1As input, after a convolution block
Obtain output valve T0；Subsequent T0Multiple RDU units containing recurrence connection are passed to, output valve T is obtainedPMDB；Since PMDB is whole
The main composition of a module is RDU unit, therefore first is unfolded to describe to RDU unit；
For convenience of understanding, the transmitting without nested RDU unit wherein information is illustrated first；With RDUkIt indicates
K-th of the RDU stacked on PMDB, wherein The total number of layer RDU unit where PMDB；Then its input value
For RDUk-1, output valve RDUk+1；In RDUkInside compresses the dimension of input value using a convolution block, is made an uproar
Acoustic signature RDUk,b1, symbol b1 is for identifying the convolution block, and and then, which is input to a duration memory intensive
Network Ndense, export RDUk,nd, symbol nd in dense network, used 8 convolution blocks, often for identifying the dense network
The characteristic spectrum number of a convolution block output is 64, therefore when finally combining 8 convolution blocks in a manner of channel attached, it obtains
To 64 × 8=512 characteristic spectrum；It is easy to cause memory overload since characteristic spectrum dimension is excessively high, therefore in dense network
Port number is dropped to 64 from 512 using 1 × 1 convolution kernel by end；Duration memory intensive network NdenseWith mathematical notation are as follows:
RDUk,nd=Hl([d1,…,di,…,d8])
Wherein HlFor NdenseThe channel reduction process of end, diIndicate NdenseI-th of the convolution block RDU in insidek,diOutput
Value, symbol [] indicate three-dimensional union operation Concatenation, finally, being connected using a global residual error by noise characteristic
It extracts, mathematical expression are as follows: RDUk+1=RDUk-1-RDUk,nd=fRDU(RDUk-1), wherein fRDUIndicate that RDU unit institute is right
The mapping function answered；
In the PMDB with nesting RDU unit, in addition to duration memory intensive network NdenseInterior roll block RDUk,di's
Other than input value is different, with the convolution block number, channel attached mode and characteristic spectrum drop used inside nesting RDU unit
Dimension is all consistent with without nested RDU unit；Dense network N inside non-nested RDU unitdense, exporting result is
diIts input value of convolution block be di-1, and come in the input value with the dense network inside nesting RDU unit, each convolution block
From in another RDU unit, i.e. output result is diIts input value of convolution block be fRDU(di-1)；
In the end duration memory intensive network module PMDB, the mode that is connected using recurrence by the output of preceding layers into
Row is three-dimensional to merge stacking, obtains the output valve T of the modulePMDB, three-dimensional union operation expression therein are as follows: TPMDB=[RDU1,…,
RDUk,…,RDUD], RDU thereinkIndicate k-th of RDU unit；
Step2.3, internal layer nested networks module carry out the fining of noise information: internal layer nested networks module is with duration
The output valve T of memory intensive network modulePMDBIt is internal that noise information is made into one using two parallel sub-networks as input
The fining of step is extracted；Image original feature channel is dropped to 64 using the filter that size is 1 × 1 by one of network path
A, another access also first reduces dimension using same method, is then reduced to 64 channels using 3 × 3 filter
32；Nested networks module can introduce more complicated Nonlinear Mapping, module end in the case where not increasing multi-parameter to network
The characteristic spectrum of two paths is carried out three-dimensional merging, the output result T obtained at this time by endninTo be added in former clear image
Noise information；
Step2.4, residual error evaluation module: the module is mainly by the resulting feature T of shallow-layer network abstraction-1Net nested with internal layer
The output T of network moduleninSubtracted each other and convolution block is used to finely tune difference, obtains estimating denoising block yest。
Further, the step Step3 includes:
Denoising block y is estimated using what network exportedestWith the known clear block y without noisenfCarry out the calculating of penalty values；
Specifically distance between the two, mathematical expression are measured using square mean error amount are as follows: The wherein total number with noise degeneration block x that N is inputted by an iteration, yest (i)Indicate that i-th band noise is degenerated
Block estimates denoising block as a result, y what is obtained after network trainingnf (i)Then for corresponding to i-th input tape noise degeneration block
Without the clear block of noise, Θ is the general designation of all parameters in Step2.
Further, the particular content of step Step4, Step5 is as follows:
Network weight is updated using Adam operator, the stop condition of circulation step Step2-Step3, iteration is,
Network losses value changes between adjacent the number of iterations is less than 1e-3；Or the number of iterations of the network in convergence reaches 200 times, repeatedly
After the completion of generation, trained network model is obtained.
The beneficial effects of the present invention are:
1, characteristic extraction procedure has used duration memory intensive network module, and the noise spot for obtaining extraction is more smart
Really；
2, iteration block has been used to connect in duration memory intensive network module, the different block energy of degree affected by noise
It is handled by uniformly combining；
3, the noise characteristic extraction later period joined internal layer nested networks module, which helps to be added more in network internal
Complicated mapping, ensure that and be unlikely to destroy the structural information of original image too much while extracting noise spot；
4, residual error evaluation module, which is added, in network end-point can largely mitigate the difficulty of network training；
5, by different training sets, which can be used for image deblocking, deblurring process, and up-sampling is added further
After layer, image super-resolution rebuilding process can be applied to；
6, the present invention can be handled in natural image in the case where not destroying original texture image marginal information with degree of precision
Existing white Gaussian noise.
Detailed description of the invention
Fig. 1 is the flow chart of step Step2 of the present invention；
The convergence graph and various algorithms corresponding PSNR value of model training when Fig. 2 is white Gaussian noise standard deviation sigma=45；
Fig. 3 is many algorithms on the picture " img004 " of standard testing collection Set12, when standard deviation sigma=45, denoises effect
And partial enlarged view, wherein (3a) original image；(3b) noisy image；(3c) EPLL algorithm；(3d) BM3D algorithm；(3e)NCSR
Algorithm；(3f) DnCNN algorithm；(3g) MemNet algorithm；DBSR algorithm (3h) of the invention；
Fig. 4 is many algorithms on the picture " img011 " of standard testing collection Set12, when standard deviation sigma=15, denoises effect
And partial enlarged view, wherein (4a) original image；(4b) noisy image；(4c) EPLL algorithm；(4d) BM3D algorithm；(4e)NCSR
Algorithm；(4f) DnCNN algorithm；(4g) MemNet algorithm；DBSR algorithm (4h) of the invention.
Specific embodiment
Embodiment 1: as shown in Figs 1-4, a kind of Efficient image denoising method based on duration memory intensive network, including
Following steps:
Step1, training set, verifying collection and test set are chosen, and training set and verifying collection is pre-processed, obtained band and make an uproar
The degeneration block of sound and the clear block without noise；
Step2, by shallow-layer feature extraction module, duration memory intensive network module, internal layer nested networks module and
Residual error evaluation module handles the degeneration block with noise, obtains estimating denoising block；
Preliminary noise feature extraction is carried out to input picture using shallow-layer feature extraction module first, will then extract gained
Information be transmitted to duration memory intensive network module, carry out fine noise using the stacking of residual error dense cell by the module
The extraction of feature；After the three-dimensional arrangement of multiple residual error dense cells output, gained information is further transferred to two accesses
Internal layer nested networks module carries out feature and refines process；Obtained information is the making an uproar of obtaining of extraction for being network at this time
Sound subtract each other can be obtained and estimates denoising by the output of shallow-layer feature extraction module and the output valve of internal layer nested networks module
Block；
Step3, using network export estimate denoising block and the known clear block without noise calculates mean square error, with into
Row loss value metric；
Step4, since network end-point, seek weight gradient value using Adam operator and update network parameter；
Step5, circulation step Step2-Step4, training terminates and obtains denoising model after network convergence；
Step6, the image of removal noise is obtained containing gaussian noise image in the input of trained denoising model.
Further, in the step Step1,900 2K used in image restoration and enhancing match NTIRE are selected
Training set, verifying collection of the high definition png image as model, wherein preceding 800 images are used for model training, latter 100 for testing
Card；It selects image denoising standard data set Set12, BSD68 to test network model, is concentrated in training set and verifying, it is right
Every 2K high definition original image adds white Gaussian noise, is then intercepted respectively in high definition original image and the same position of noise pattern with step-length 27
The image block that size is 96 × 96, obtains the clear block y without noisenfWith the degeneration block x with noise.
Further, in the step 2, the network first floor is preliminary using the shallow-layer feature extraction module comprising single convolution block
Extract noise characteristic；Then, it using residual error dense cell as basic unit, and is held by the layer-by-layer nesting of basic unit to construct
Continuous property memory intensive network module, to extract more granularity noise characteristic information；In the latter half of network, internal layer nested networks make
Channel integration is carried out to noise characteristic map with two parallel sub-networks, is realized by different numbers, different size of convolution kernel
The fining of noise information feature, to obtain image high-frequency noise information；Finally, network end-point will include high-frequency noise information
Picture material is integrated with the noise characteristic tentatively extracted, by asking the way of difference to remove noise in residual error evaluation module
Image.
Further, the specific steps of the step Step2:
Step2.1, shallow-layer feature extraction module tentatively extract noise characteristic: using the degeneration block x with noise as input value,
Increase the port number of input block by shallow-layer feature extraction module, and carries out the preliminary extraction of noise information to it；
Shallow-layer feature extraction module using the degeneration block x with noise as input, by convolution operation Conv, weight bias and
After band parameters revision linear unit PReLU, T is exported-1；The process is expressed with mathematical formulae are as follows: T-1=PReLU (aC1,C1),
Middle C1=W1*x+b1The output of process, W are biased for convolution1For the weight parameter of convolution operation, b1For amount of bias, a is that band parameter is repaired
The initial value of linear positive unit PReLU；
Step2.2, duration memory intensive network module refine noise characteristic information:
Duration memory intensive network module (persistent memory dense block, PMDB) is intensive with residual error
Unit (Residual Dense Unit, RDU) is used as its basic unit, passes through the way of nesting RDU further inside RDU
To achieve the purpose that extract more granularity noise characteristic information；
Specific practice are as follows: PMDB is with the output valve T of shallow-layer feature extraction module-1As input, after a convolution block
Obtain output valve T0；Subsequent T0Multiple RDU units containing recurrence connection are passed to, output valve T is obtainedPMDB；Since PMDB is whole
The main composition of a module is RDU unit, therefore first is unfolded to describe to RDU unit；
For convenience of understanding, the transmitting without nested RDU unit wherein information is illustrated first；With RDUkIt indicates
K-th of the RDU stacked on PMDB, wherein The total number of layer RDU unit where PMDB；Then its input value
For RDUk-1, output valve RDUk+1；In RDUkInside compresses the dimension of input value using a convolution block, is made an uproar
Acoustic signature RDUk,b1, symbol b1 is for identifying the convolution block, and and then, which is input to a duration memory intensive
Network Ndense, export RDUk,nd, symbol nd in dense network, used 8 convolution blocks, often for identifying the dense network
The characteristic spectrum number of a convolution block output is 64, therefore when finally combining 8 convolution blocks in a manner of channel attached, it obtains
To 64 × 8=512 characteristic spectrum；It is easy to cause memory overload since characteristic spectrum dimension is excessively high, therefore in dense network
Port number is dropped to 64 from 512 using 1 × 1 convolution kernel by end；Duration memory intensive network NdenseWith mathematical notation are as follows:
RDUk,nd=Hl([d1,…,di,…,d8])
Wherein HlFor NdenseThe channel reduction process of end, diIndicate NdenseI-th of the convolution block RDU in insidek,diOutput
Value, symbol [] indicate three-dimensional union operation Concatenation, finally, being connected using a global residual error by noise characteristic
It extracts, mathematical expression are as follows: RDUk+1=RDUk-1-RDUk,nd=fRDU(RDUk-1), wherein fRDUIndicate that RDU unit institute is right
The mapping function answered；
In the PMDB with nesting RDU unit, in addition to duration memory intensive network NdenseInterior roll block RDUk,di's
Other than input value is different, with the convolution block number, channel attached mode and characteristic spectrum drop used inside nesting RDU unit
Dimension is all consistent with without nested RDU unit；Dense network N inside non-nested RDU unitdense, exporting result is
diIts input value of convolution block be di-1, and come in the input value with the dense network inside nesting RDU unit, each convolution block
From in another RDU unit, i.e. output result is diIts input value of convolution block be fRDU(di-1)；
In the end duration memory intensive network module PMDB, the mode that is connected using recurrence by the output of preceding layers into
Row is three-dimensional to merge stacking, obtains the output valve T of the modulePMDB, three-dimensional union operation expression therein are as follows: TPMDB=[RDU1,…,
RDUk,…,RDUD], RDU thereinkIndicate k-th of RDU unit；
Step2.3, internal layer nested networks module carry out the fining of noise information: internal layer nested networks module is with duration
The output valve T of memory intensive network modulePMDBIt is internal that noise information is made into one using two parallel sub-networks as input
The fining of step is extracted；Image original feature channel is dropped to 64 using the filter that size is 1 × 1 by one of network path
A, another access also first reduces dimension using same method, is then reduced to 64 channels using 3 × 3 filter
32；Nested networks module can introduce more complicated Nonlinear Mapping, module end in the case where not increasing multi-parameter to network
The characteristic spectrum of two paths is carried out three-dimensional merging, the output result T obtained at this time by endninTo be added in former clear image
Noise information；
Step2.4, residual error evaluation module: the module is mainly by the resulting feature T of shallow-layer network abstraction-1Net nested with internal layer
The output T of network moduleninSubtracted each other and convolution block is used to finely tune difference, obtains estimating denoising block yest。
Further, the step Step3 includes:
Denoising block y is estimated using what network exportedestWith the known clear block y without noisenfCarry out the calculating of penalty values；
Specifically distance between the two, mathematical expression are measured using square mean error amount are as follows: The wherein total number with noise degeneration block x that N is inputted by an iteration, yest (i)Indicate that i-th band noise is degenerated
Block estimates denoising block as a result, y what is obtained after network trainingnf (i)Then for corresponding to i-th input tape noise degeneration block
Without the clear block of noise, Θ is the general designation of all parameters in Step2.
Network training parameter of the invention and training objective are described as follows:
The training parameter for including in Step2 has: single convolution block, duration memory intensive net in shallow-layer feature extraction module
Convolution kernel size included in network module PMDB and internal layer nested networks module, convolution kernel number, band parameters revision are linear
The initial value of unit；Learning rate involved in Adam operator in Step4, the number of iterations etc. of iterative process.
Training objective of the invention is to have converged in Step3 in the penalty values that model is completed to be calculated when training
One lower value.
Further, the particular content of step Step4, Step5 is as follows:
Network weight is updated using Adam operator, the stop condition of circulation step Step2-Step3, iteration is,
Network losses value changes between adjacent the number of iterations is less than 1e-3；Or the number of iterations of the network in convergence reaches 200 times, repeatedly
After the completion of generation, trained network model is obtained.
In order to illustrate effect of the invention, the convergence of model training when being illustrated in figure 2 white Gaussian noise standard deviation sigma=45
Figure and the corresponding PSNR value of various algorithms；
As can be seen from Figure 2, Denoising Algorithm proposed by the invention has greater advantage on Y-PSNR PSNR, compared to
Its promotion amplitude of the MemNet being recently proposed is 0.11dB, and compared to traditional classical Denoising Algorithm BM3D, promotion amplitude is more apparent,
Size is 1.54dB.
Fig. 3 is qualitative to compare several classic algorithm, while removing noise to the fidelity of original image detailed information.
Comparison discovery, still there is more noise in the restored image of classic algorithm such as BM3D, NCSR, and the algorithm based on deep learning
The noise of DnCNN/MemNet restored image has been efficiently removed, but detailed information is distorted.The present invention removal noise with
To being significantly better than that aforementioned algorism in terms of the fidelity situation two of prime information.
Fig. 4 is qualitative to compare several classic algorithm, right in the higher situation of fidelity to original image detailed information
The ability of noise remove.Noise on comparison discovery, classic algorithm such as BM3D, NCSR and algorithm DnCNN based on deep learning
Information is still more, and noise image its noise that MemNet and the present invention are handled has been effectively suppressed, but the present invention is believing
It is better than MemNet algorithm in terms of breath fidelity.
Above in conjunction with attached drawing, the embodiment of the present invention is explained in detail, but the present invention is not limited to above-mentioned
Embodiment within the knowledge of a person skilled in the art can also be before not departing from present inventive concept
Put that various changes can be made.
Claims (6)
1. a kind of Efficient image denoising method based on duration memory intensive network, characterized by the following steps:
Step1, training set, verifying collection and test set are chosen, and training set and verifying collection is pre-processed, obtained with noise
Degeneration block and the clear block without noise；
Step2, pass through shallow-layer feature extraction module, duration memory intensive network module, internal layer nested networks module and residual error
Evaluation module handles the degeneration block with noise, obtains estimating denoising block；
Preliminary noise feature extraction is carried out to input picture using shallow-layer feature extraction module first, will then extract resulting letter
Breath is transmitted to duration memory intensive network module, carries out fine noise characteristic using the stacking of residual error dense cell by the module
Extraction；After the three-dimensional arrangement of multiple residual error dense cells output, gained information is further transferred to the internal layer of two accesses
Nested networks module carries out feature and refines process；Obtained information is the obtained noise of extraction for being network at this time, by
The output of shallow-layer feature extraction module and the output valve of internal layer nested networks module, which subtract each other can be obtained, estimates denoising block；
Step3, denoising block and the known clear block calculating mean square error without noise are estimated using what network exported, to be damaged
Lose value metric；
Step4, since network end-point, seek weight gradient value using Adam operator and update network parameter；
Step5, circulation step Step2-Step4, training terminates and obtains denoising model after network convergence；
Step6, the image of removal noise is obtained containing gaussian noise image in the input of trained denoising model.
2. the Efficient image denoising method according to claim 1 based on duration memory intensive network, it is characterised in that:
In the step Step1, select 900 2K high definition png images used in image restoration and enhancing match NTIRE as mould
Training set, the verifying collection of type, wherein preceding 800 images are used for model training, latter 100 for verifying；Select image denoising mark
Quasi- data set Set12, BSD68 test network model, concentrate in training set and verifying, add to every 2K high definition original image
White Gaussian noise, the image for being respectively then 96 × 96 in high definition original image and the same position of noise pattern interception size with step-length 27
Block obtains the clear block y without noisenfWith the degeneration block x with noise.
3. the Efficient image denoising method according to claim 1 based on duration memory intensive network, it is characterised in that:
In the step 2, the network first floor tentatively extracts noise characteristic using the shallow-layer feature extraction module comprising single convolution block；With
Afterwards, using residual error dense cell as basic unit, and duration memory intensive net is constructed by the layer-by-layer nesting of basic unit
Network module, to extract more granularity noise characteristic information；In the latter half of network, internal layer nested networks use two parallel subnets
Network carries out channel integration to noise characteristic map, realizes noise information feature by different numbers, different size of convolution kernel
Fining, to obtain image high-frequency noise information；Finally, network end-point is by the picture material comprising high-frequency noise information and tentatively
The noise characteristic of extraction is integrated, and removes the image of noise by seeking the way of difference in residual error evaluation module.
4. the Efficient image denoising method according to claim 1 based on duration memory intensive network, it is characterised in that:
The specific steps of the step Step2:
Step2.1, shallow-layer feature extraction module tentatively extract noise characteristic: using the degeneration block x with noise as input value, passing through
Shallow-layer feature extraction module increases the port number of input block, and the preliminary extraction of noise information is carried out to it；
Shallow-layer feature extraction module is using the degeneration block x with noise as input, by convolution operation Conv, weight bias and with ginseng
After number amendment linear unit PReLU, T is exported-1；The process is expressed with mathematical formulae are as follows: T-1=PReLU (aC1,C1), wherein C1
=W1*x+b1The output of process, W are biased for convolution1For the weight parameter of convolution operation, b1For amount of bias, a is band parameters revision
The initial value of linear unit PReLU；
Step2.2, duration memory intensive network module refine noise characteristic information:
Duration memory intensive network module (persistent memory dense block, PMDB) is with residual error dense cell
(Residual Dense Unit, RDU) is used as its basic unit, is reached by the way of nesting RDU further inside RDU
To the purpose for extracting more granularity noise characteristic information；
Specific practice are as follows: PMDB is with the output valve T of shallow-layer feature extraction module-1As input, obtained after a convolution block
Output valve T0；Subsequent T0Multiple RDU units containing recurrence connection are passed to, output valve T is obtainedPMDB；Due to the entire mould of PMDB
The main composition of block is RDU unit, therefore first is unfolded to describe to RDU unit；
For convenience of understanding, the transmitting without nested RDU unit wherein information is illustrated first；With RDUkIt indicates on PMDB
K-th of the RDU stacked, wherein The total number of layer RDU unit where PMDB；Then its input value is
RDUk-1, output valve RDUk+1；In RDUkInside compresses the dimension of input value using a convolution block, obtains noise
Feature RDUk,b1, symbol b1 is for identifying the convolution block, and and then, which is input to a duration memory intensive net
Network Ndense, export RDUk,nd, symbol nd in dense network, used 8 convolution blocks, each for identifying the dense network
The characteristic spectrum number of convolution block output is 64, therefore when finally combining 8 convolution blocks in a manner of channel attached, it obtains
64 × 8=512 characteristic spectrum；Memory overload is easy to cause since characteristic spectrum dimension is excessively high, therefore at dense network end
Port number is dropped to 64 from 512 using 1 × 1 convolution kernel by end；Duration memory intensive network NdenseWith mathematical notation are as follows:
RDUk,nd=Hl([d1,…,di,…,d8])
Wherein HlFor NdenseThe channel reduction process of end, diIndicate NdenseI-th of the convolution block RDU in insidek,diOutput valve, symbol
Number [] indicates three-dimensional union operation Concatenation, finally, being extracted noise characteristic using a global residual error connection
Come, mathematical expression are as follows: RDUk+1=RDUk-1-RDUk,nd=fRDU(RDUk-1), wherein fRDUIt indicates to reflect corresponding to the RDU unit
Penetrate function；
In the PMDB with nesting RDU unit, in addition to duration memory intensive network NdenseInterior roll block RDUk,diInput
Other than value is different, all with convolution block number, channel attached mode and the characteristic spectrum dimensionality reduction used inside nesting RDU unit
It is consistent with without nested RDU unit；Dense network N inside non-nested RDU unitdense, output result is di's
Its input value of convolution block is di-1, and with the dense network inside nesting RDU unit, the input value of each convolution block from
Another RDU unit, i.e. output result are diIts input value of convolution block be fRDU(di-1)；
In the end duration memory intensive network module PMDB, the output of preceding layers is carried out three using the mode that recurrence connects
Dimension, which merges, to be stacked, and the output valve T of the module is obtainedPMDB, three-dimensional union operation expression therein are as follows: TPMDB=[RDU1,…,
RDUk,…,RDUD], RDU thereinkIndicate k-th of RDU unit；
Step2.3, internal layer nested networks module carry out the fining of noise information: internal layer nested networks module is remembered with duration
The output valve T of dense network modulePMDBIt is internal that noise information is made further using two parallel sub-networks as input
Fining is extracted；Image original feature channel is dropped to 64 using the filter that size is 1 × 1 by one of network path,
Another access also first reduces dimension using same method, and 64 channels are then reduced to 32 using 3 × 3 filter；
Nested networks module can introduce more complicated Nonlinear Mapping in the case where not increasing multi-parameter to network, and end of module will
The characteristic spectrum of two paths carries out three-dimensional merging, the output result T obtained at this timeninIt makes an uproar for what is be added in former clear image
Acoustic intelligence；
Step2.4, residual error evaluation module: the module is mainly by the resulting feature T of shallow-layer network abstraction-1With internal layer nested networks mould
The output T of blockninSubtracted each other and convolution block is used to finely tune difference, obtains estimating denoising block yest。
5. the Efficient image denoising method according to claim 4 based on duration memory intensive network, it is characterised in that:
The step Step3 includes:
Denoising block y is estimated using what network exportedestWith the known clear block y without noisenfCarry out the calculating of penalty values；Specifically
It is the distance measured using square mean error amount between the two, mathematical expression are as follows: Its
The total number with noise degeneration block x that middle N is inputted by an iteration, yest (i)Indicate i-th band noise degeneration block by net
What is obtained after network training estimates denoising block as a result, ynf (i)It is then corresponding to i-th input tape noise degeneration block without noise
Clear block, Θ are the general designation of all parameters in Step2.
6. the Efficient image denoising method according to claim 4 based on duration memory intensive network, it is characterised in that:
The particular content of step Step4, Step5 is as follows:
Network weight is updated using Adam operator, the stop condition of circulation step Step2-Step3, iteration are networks
Penalty values change between adjacent the number of iterations is less than 1e-3；Or the number of iterations of the network in convergence reaches 200 times, iteration is complete
Cheng Hou obtains trained network model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811576192.5A CN109785249A (en) | 2018-12-22 | 2018-12-22 | A kind of Efficient image denoising method based on duration memory intensive network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811576192.5A CN109785249A (en) | 2018-12-22 | 2018-12-22 | A kind of Efficient image denoising method based on duration memory intensive network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109785249A true CN109785249A (en) | 2019-05-21 |
Family
ID=66497556
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811576192.5A Pending CN109785249A (en) | 2018-12-22 | 2018-12-22 | A kind of Efficient image denoising method based on duration memory intensive network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109785249A (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110232361A (en) * | 2019-06-18 | 2019-09-13 | 中国科学院合肥物质科学研究院 | Human body behavior intension recognizing method and system based on the dense network of three-dimensional residual error |
| CN110349103A (en) * | 2019-07-01 | 2019-10-18 | 昆明理工大学 | It is a kind of based on deep neural network and jump connection without clean label image denoising method |
| CN110490816A (en) * | 2019-07-15 | 2019-11-22 | 哈尔滨工程大学 | A kind of underwater Heterogeneous Information data noise reduction |
| CN110569738A (en) * | 2019-08-15 | 2019-12-13 | 杨春立 | natural scene text detection method, equipment and medium based on dense connection network |
| CN110738231A (en) * | 2019-07-25 | 2020-01-31 | 太原理工大学 | Method for classifying mammary gland X-ray images by improving S-DNet neural network model |
| CN110838095A (en) * | 2019-11-06 | 2020-02-25 | 广西师范大学 | Single image rain removing method and system based on cyclic dense neural network |
| CN111105375A (en) * | 2019-12-17 | 2020-05-05 | 北京金山云网络技术有限公司 | Image generation method, model training method and device thereof, and electronic equipment |
| CN111275643A (en) * | 2020-01-20 | 2020-06-12 | 西南科技大学 | True noise blind denoising network model and method based on channel and space attention |
| CN111681298A (en) * | 2020-06-08 | 2020-09-18 | 南开大学 | Compressed sensing image reconstruction method based on multi-feature residual error network |
| CN112150384A (en) * | 2020-09-29 | 2020-12-29 | 中科方寸知微(南京)科技有限公司 | Method and system based on fusion of residual error network and dynamic convolution network model |
| CN112419219A (en) * | 2020-11-25 | 2021-02-26 | 广州虎牙科技有限公司 | Image enhancement model training method, image enhancement method and related device |
| CN113284059A (en) * | 2021-04-29 | 2021-08-20 | Oppo广东移动通信有限公司 | Model training method, image enhancement method, device, electronic device and medium |
| EP3985972A4 (en) * | 2019-10-16 | 2022-11-16 | Tencent Technology (Shenzhen) Company Limited | Machine learning-based artifact removal method and apparatus, and machine learning-based artifact removal model training method and apparatus |
| CN116051408A (en) * | 2023-01-06 | 2023-05-02 | 郑州轻工业大学 | Image depth denoising method based on residual error self-coding |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102156960A (en) * | 2010-12-16 | 2011-08-17 | 新奥特（北京）视频技术有限公司 | Picture noise adding method |
| US20150086126A1 (en) * | 2012-04-27 | 2015-03-26 | Nec Corporation | Image processing method, image processing system, image processing device, and image processing device |
| US20150339806A1 (en) * | 2014-05-26 | 2015-11-26 | Fujitsu Limited | Image denoising method and image denoising apparatus |
| CN106709875A (en) * | 2016-12-30 | 2017-05-24 | 北京工业大学 | Compressed low-resolution image restoration method based on combined deep network |
| CN108304755A (en) * | 2017-03-08 | 2018-07-20 | 腾讯科技（深圳）有限公司 | The training method and device of neural network model for image procossing |
| CN108492265A (en) * | 2018-03-16 | 2018-09-04 | 西安电子科技大学 | CFA image demosaicing based on GAN combines denoising method |
| CN108629736A (en) * | 2017-03-15 | 2018-10-09 | 三星电子株式会社 | System and method for designing super-resolution depth convolutional neural networks |
| CN108961186A (en) * | 2018-06-29 | 2018-12-07 | 赵岩 | A kind of old film reparation recasting method based on deep learning |
| CN109767386A (en) * | 2018-12-22 | 2019-05-17 | 昆明理工大学 | A kind of rapid image super resolution ratio reconstruction method based on deep learning |
-
2018
- 2018-12-22 CN CN201811576192.5A patent/CN109785249A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102156960A (en) * | 2010-12-16 | 2011-08-17 | 新奥特（北京）视频技术有限公司 | Picture noise adding method |
| US20150086126A1 (en) * | 2012-04-27 | 2015-03-26 | Nec Corporation | Image processing method, image processing system, image processing device, and image processing device |
| US20150339806A1 (en) * | 2014-05-26 | 2015-11-26 | Fujitsu Limited | Image denoising method and image denoising apparatus |
| CN106709875A (en) * | 2016-12-30 | 2017-05-24 | 北京工业大学 | Compressed low-resolution image restoration method based on combined deep network |
| CN108304755A (en) * | 2017-03-08 | 2018-07-20 | 腾讯科技（深圳）有限公司 | The training method and device of neural network model for image procossing |
| CN108629736A (en) * | 2017-03-15 | 2018-10-09 | 三星电子株式会社 | System and method for designing super-resolution depth convolutional neural networks |
| CN108492265A (en) * | 2018-03-16 | 2018-09-04 | 西安电子科技大学 | CFA image demosaicing based on GAN combines denoising method |
| CN108961186A (en) * | 2018-06-29 | 2018-12-07 | 赵岩 | A kind of old film reparation recasting method based on deep learning |
| CN109767386A (en) * | 2018-12-22 | 2019-05-17 | 昆明理工大学 | A kind of rapid image super resolution ratio reconstruction method based on deep learning |
Non-Patent Citations (4)
| Title |
|---|
| BAOSHUN SHI等: "SBM3D: Sparse Regularization Model Induced by BM3D for Weighted Diffraction Imaging", 《IEEE ACCESS》 * |
| GIORGIO PATRINI等: "Making Deep Neural Networks Robust to Label Noise: A Loss Correction Approach", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》 * |
| 刘辉等: "基于均匀空间色差度量的彩色图像椒盐噪声滤波算法", 《传感器与微系统》 * |
| 高净植等: "改进深度残差卷积神经网络的LDCT图像估计", 《计算机工程与应用》 * |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110232361B (en) * | 2019-06-18 | 2021-04-02 | 中国科学院合肥物质科学研究院 | Human behavior intention identification method and system based on three-dimensional residual dense network |
| CN110232361A (en) * | 2019-06-18 | 2019-09-13 | 中国科学院合肥物质科学研究院 | Human body behavior intension recognizing method and system based on the dense network of three-dimensional residual error |
| CN110349103A (en) * | 2019-07-01 | 2019-10-18 | 昆明理工大学 | It is a kind of based on deep neural network and jump connection without clean label image denoising method |
| CN110490816A (en) * | 2019-07-15 | 2019-11-22 | 哈尔滨工程大学 | A kind of underwater Heterogeneous Information data noise reduction |
| CN110490816B (en) * | 2019-07-15 | 2022-11-15 | 哈尔滨工程大学 | Underwater heterogeneous information data noise reduction method |
| CN110738231A (en) * | 2019-07-25 | 2020-01-31 | 太原理工大学 | Method for classifying mammary gland X-ray images by improving S-DNet neural network model |
| CN110738231B (en) * | 2019-07-25 | 2022-12-27 | 太原理工大学 | Method for classifying mammary gland X-ray images by improving S-DNet neural network model |
| CN110569738A (en) * | 2019-08-15 | 2019-12-13 | 杨春立 | natural scene text detection method, equipment and medium based on dense connection network |
| CN110569738B (en) * | 2019-08-15 | 2023-06-06 | 杨春立 | Natural scene text detection method, equipment and medium based on densely connected network |
| US11985358B2 (en) | 2019-10-16 | 2024-05-14 | Tencent Technology (Shenzhen) Company Limited | Artifact removal method and apparatus based on machine learning, and method and apparatus for training artifact removal model based on machine learning |
| EP3985972A4 (en) * | 2019-10-16 | 2022-11-16 | Tencent Technology (Shenzhen) Company Limited | Machine learning-based artifact removal method and apparatus, and machine learning-based artifact removal model training method and apparatus |
| CN110838095A (en) * | 2019-11-06 | 2020-02-25 | 广西师范大学 | Single image rain removing method and system based on cyclic dense neural network |
| CN110838095B (en) * | 2019-11-06 | 2022-06-07 | 广西师范大学 | Single image rain removing method and system based on cyclic dense neural network |
| CN111105375A (en) * | 2019-12-17 | 2020-05-05 | 北京金山云网络技术有限公司 | Image generation method, model training method and device thereof, and electronic equipment |
| CN111105375B (en) * | 2019-12-17 | 2023-08-22 | 北京金山云网络技术有限公司 | Image generation method, model training method and device thereof, and electronic equipment |
| CN111275643B (en) * | 2020-01-20 | 2022-09-02 | 西南科技大学 | Real noise blind denoising network system and method based on channel and space attention |
| CN111275643A (en) * | 2020-01-20 | 2020-06-12 | 西南科技大学 | True noise blind denoising network model and method based on channel and space attention |
| CN111681298A (en) * | 2020-06-08 | 2020-09-18 | 南开大学 | Compressed sensing image reconstruction method based on multi-feature residual error network |
| CN112150384A (en) * | 2020-09-29 | 2020-12-29 | 中科方寸知微(南京)科技有限公司 | Method and system based on fusion of residual error network and dynamic convolution network model |
| CN112150384B (en) * | 2020-09-29 | 2024-03-29 | 中科方寸知微(南京)科技有限公司 | Method and system based on fusion of residual network and dynamic convolution network model |
| CN112419219A (en) * | 2020-11-25 | 2021-02-26 | 广州虎牙科技有限公司 | Image enhancement model training method, image enhancement method and related device |
| CN113284059A (en) * | 2021-04-29 | 2021-08-20 | Oppo广东移动通信有限公司 | Model training method, image enhancement method, device, electronic device and medium |
| CN116051408B (en) * | 2023-01-06 | 2023-10-27 | 郑州轻工业大学 | Image depth denoising method based on residual error self-coding |
| CN116051408A (en) * | 2023-01-06 | 2023-05-02 | 郑州轻工业大学 | Image depth denoising method based on residual error self-coding |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109785249A (en) | A kind of Efficient image denoising method based on duration memory intensive network | |
| CN106204467B (en) | Image denoising method based on cascade residual error neural network | |
| CN103927531B (en) | It is a kind of based on local binary and the face identification method of particle group optimizing BP neural network | |
| CN104361328B (en) | A kind of facial image normalization method based on adaptive multiple row depth model | |
| CN110796625B (en) | Image compressed sensing reconstruction method based on group sparse representation and weighted total variation | |
| De-Maeztu et al. | Near real-time stereo matching using geodesic diffusion | |
| CN109064423B (en) | Intelligent image repairing method for generating antagonistic loss based on asymmetric circulation | |
| CN107748895A (en) | UAV Landing landforms image classification method based on DCT CNN models | |
| CN110648292A (en) | High-noise image denoising method based on deep convolutional network | |
| CN106204482A (en) | Based on the mixed noise minimizing technology that weighting is sparse | |
| CN111723701A (en) | Underwater target identification method | |
| CN114663685B (en) | Pedestrian re-recognition model training method, device and equipment | |
| CN112200733B (en) | Grid denoising method based on graph convolution network | |
| Su et al. | Multi‐scale cross‐path concatenation residual network for Poisson denoising | |
| CN114202017A (en) | SAR optical image mapping model lightweight method based on condition generation countermeasure network | |
| Yap et al. | A recursive soft-decision approach to blind image deconvolution | |
| CN105913451B (en) | A kind of natural image superpixel segmentation method based on graph model | |
| CN111291810A (en) | Information processing model generation method based on target attribute decoupling and related equipment | |
| CN114998107A (en) | Image blind super-resolution network model, method, equipment and storage medium | |
| Xu et al. | Dual-branch deep image prior for image denoising | |
| CN103037168A (en) | Stable Surfacelet domain multi-focus image fusion method based on compound type pulse coupled neural network (PCNN) | |
| CN116405100B (en) | Distortion signal restoration method based on priori knowledge | |
| Nejati et al. | Low-rank regularized collaborative filtering for image denoising | |
| Zou et al. | EDCNN: a novel network for image denoising | |
| CN111047537A (en) | System for recovering details in image denoising |
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 | ||
| AD01 | Patent right deemed abandoned | ||
| AD01 | Patent right deemed abandoned |
Effective date of abandoning: 20230707 |