CN114663312A - Network automatic searching method aiming at image noise reduction and image noise reduction method - Google Patents
Network automatic searching method aiming at image noise reduction and image noise reduction method Download PDFInfo
- Publication number
- CN114663312A CN114663312A CN202210305751.9A CN202210305751A CN114663312A CN 114663312 A CN114663312 A CN 114663312A CN 202210305751 A CN202210305751 A CN 202210305751A CN 114663312 A CN114663312 A CN 114663312A
- Authority
- CN
- China
- Prior art keywords
- noise reduction
- module
- cnn
- network
- image
- 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
- 230000009467 reduction Effects 0.000 title claims abstract description 62
- 238000000034 method Methods 0.000 title claims abstract description 51
- 230000007704 transition Effects 0.000 claims abstract description 14
- 238000012549 training Methods 0.000 claims description 13
- 230000035772 mutation Effects 0.000 claims description 4
- 238000012360 testing method Methods 0.000 claims description 4
- 238000000638 solvent extraction Methods 0.000 claims description 2
- 238000004422 calculation algorithm Methods 0.000 abstract description 37
- 230000000694 effects Effects 0.000 abstract description 15
- 238000012545 processing Methods 0.000 abstract description 12
- 238000010586 diagram Methods 0.000 abstract description 8
- 230000002068 genetic effect Effects 0.000 abstract description 5
- 238000013527 convolutional neural network Methods 0.000 description 74
- 238000013461 design Methods 0.000 description 6
- 230000006870 function Effects 0.000 description 6
- 238000003909 pattern recognition Methods 0.000 description 3
- 238000012795 verification Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000004069 differentiation Effects 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000002787 reinforcement Effects 0.000 description 2
- 238000010845 search algorithm Methods 0.000 description 2
- 241000196324 Embryophyta Species 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 235000015220 hamburgers Nutrition 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Physiology (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a network automatic searching method aiming at image noise reduction and an image noise reduction method, and belongs to the technical field of image processing. When a network structure for image noise reduction is constructed, a Feature module is used as a first layer of a CNN network structure to ensure that an input image has enough characteristics to learn a deep module, a Transition module transforms the characteristic dimensions of a characteristic diagram, and a Dropout module prevents overfitting of the network, so that a better Gaussian noise reduction effect is obtained; in addition, the automatic structure searching method based on the genetic algorithm is adopted, the automatically designed network structure can be automatically adjusted under the data set according to different noise level problems, the time cost problem of manual intervention is avoided, and the speed and the performance of processing the noise reduction task are greatly improved.
Description
Technical Field
The invention relates to a network automatic searching method aiming at image noise reduction and an image noise reduction method, belonging to the technical field of image processing.
Background
In recent years, Convolutional Neural Networks (CNNs) have achieved enormous efforts in dealing with computer vision tasks, and have become the mainstream methods for solving computer vision problems such as image classification, target detection, and target tracking. The network is deeper from the initial superposition of the network layer number, the residual error structure and the dense connection structure of the network layer features with different depths are fused later, and the structure of the network width feature fusion is further formed, for example, a residual error network structure ResNet, a dense connection network structure DenseNet and a GoogleNet (inclusion) structure of the width feature fusion are respectively shown in figures 1, 2 and 3. Obviously, the CNN network structure has become a key to influence the performance of CNN models.
The image denoising refers to a process of reducing noise in a digital image, the digital image in reality is often affected by noise interference of an imaging device and an external environment in the digitization and transmission processes, and is called a noisy image or a noise image, and the presence of the noise affects subsequent further processing of the image, such as classifying the image, performing target recognition and tracking, and the like.
For the image noise reduction task, a CNN network for manually designing the image noise reduction task is provided, and comprises a BM3D algorithm and a WNNM algorithm based on a nonlinear similarity principle, an EPLL algorithm based on a generation model, an MLP algorithm based on an independent training principle, a CSF algorithm and a TNRD algorithm, and a DnCNN algorithm. However, the conventional manually designed image denoising CNN network structure not only needs a lot of professional knowledge and experience related to CNN, but also needs to design different CNN network structures for different data sets, which greatly limits the development of the image denoising CNN structure.
In order to solve the above problems, algorithms capable of automatically designing a CNN network structure are proposed, which are capable of automatically searching for a CNN structure having superior performance without requiring artificial knowledge and experience. These algorithms can be broadly classified into two broad categories: semi-automatic CNN structure search algorithms represented by algorithms such as Genetic CNN, historical Evolution, EAS, Block-QNN-S and the like, and full-automatic CNN structure search algorithms represented by Large-scale Evolution, CGP-CNN, NAS, MetaQNN, IPPSO, AE-CNN and the like. It should be noted that both of these algorithms are based on evolutionary computing algorithm or reinforcement learning algorithm, such as Genetic CNN, historical Evolution, CGP-CNN, Large-scale Evolution, AE-CNN, IIPSO is based on evolutionary computing algorithm, NSA, MetaQN, EAS and Block-QNN-S are based on reinforcement learning algorithm. However, the algorithms are all used for solving the image classification task, the last module of the network structure corresponding to the image classification task is a classifier, and one-dimensional data is output to represent the image classification result; and the image noise reduction is carried out, the corresponding network structure is finally a convolution module, and the output result is the noise of the two-dimensional data representation image, so the problem of image noise reduction cannot be solved by directly adopting the existing network structure automatic search method aiming at image classification.
Moreover, the CNN structure design for image noise reduction, especially for gaussian noise reduction task, is an optimization problem, and the loss function L (M, T) is minimized on the training data set T by optimizing the network model M. The CNN model M can be represented by a network internal structure coding vector λ and is learned by a specific learning algorithm a on a training set T, i.e., M ═ a (λ, T). The design goal of the CNN structure is to find an optimal network structure solution lambda*So that M*＝A(λ*T), and L (M) is satisfied on the verification set V*V) minimum, i.e. expressed by the following mathematical formula:
λ*＝argλminL(T,M)＝argλminf(λ,A,T,V,L) (1)
the target function f receives a vector lambda representing network structure coding as input, outputs a corresponding loss value, and a training set T and a verification set V are divided by the training set according to a certain proportion to meet the requirement
By minimizing the loss value of the objective function f, an optimal λ can be obtained, and then a network model M can be constructed. And finally, verifying the generalization ability of M in the test set. Theoretically, the network structure λ can be optimized by gradient descent and other algorithms, but it is difficult to implement in practice for the following reasons:
1. the essential of optimizing the loss value of the objective function f is to find a specific lambda, which belongs to the problem of combinatorial optimization, i.e. the noise reduction index values of all network structures need to be calculated, and the calculation cost is high.
2. The structure vector λ representing the network model M is a discrete code that cannot be handled by conventional methods of handling continuous functions.
Disclosure of Invention
In order to better solve the problem of image noise reduction, the invention provides an automatic network searching method and an image noise reduction method aiming at image noise reduction, aiming at the problem that the traditional network structure based on a CNN processing Gaussian noise reduction task is only a simple linear structure and greatly limits the effect of Gaussian noise reduction, the application provides a depth structure based on ResNet and DenseNet and an inclusion width structure, and simultaneously introduces a Feature extraction module, a Transition Feature layer transformation module and a Dropout noise reduction module, fully utilizes the depth and width structure of the CNN network structure, and greatly improves the effect of the CNN network processing Gaussian noise reduction task.
A method for network auto-search for image noise reduction, the method comprising:
step 2, dividing the public image noise reduction data set according to a preset ratio, and calculating the fitness value of each CNN network individual in the population on the divided data set;
step 3, performing crossover and mutation operation according to the fitness value of each CNN network individual to generate filial generations;
step 4, generating a new population P by natural selection of CNN network individuals of the offspring and the father;
and 5, repeating the steps 2-4 until a termination condition is met, wherein the finally obtained CNN network individual is a searched network structure for image noise reduction.
Optionally, the calculating, in step 2, the fitness value of each CNN network individual in the population on the partitioned data sets includes:
and (3) training each CNN network individual on a divided data set by 30 epochs to obtain a peak signal-to-noise ratio (PSNR) as a fitness value of the corresponding CNN network individual.
Optionally, the Feature module comprises two branches, i.e., ConvRelu1, ConvRelu2, and ConvRelu 3; ConvRelu3 corresponds to a convolution kernel size of 3 x 3, Padding of 1, move step Stride of 1, Output Channel of c, Output Channel of 32.
Optionally, the Transition module takes Conv, BN and ReLU operations, where Conv has a convolution kernel size of 1 × 1, Padding of 0, and a move step Stride of 1.
Optionally, the image denoising data Set disclosed in step 2 is a Set12 and BM3D data Set.
Optionally, the predetermined ratio is 1: 5.
Optionally, before calculating the fitness value of each CNN network individual in the population on the partitioned data set, the method further includes partitioning the test set and the training set according to a ratio of 1:4 on the partitioned data set.
The application also provides an image noise reduction method, and the method adopts the network structure for image noise reduction searched by the method to perform image noise reduction processing.
The invention has the beneficial effects that:
the method has the advantages that various structural information of the CNN can be fully utilized through the introduced Resnet and DenseNet depth module, the inclusion width module, the Feature characteristic learning module, the Dropout module and the Transition module, and the problem of Gaussian noise reduction is solved better; in addition, the automatic structure searching method based on the genetic algorithm is adopted, the automatically designed network structure can be automatically adjusted under the data set according to different noise level problems, the time cost problem of manual intervention is avoided, and the speed and the performance of processing the noise reduction task are greatly improved.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
Fig. 1 is a structural diagram of a residual error network ResNet in a CNN network.
Fig. 2 is a schematic structural diagram of a dense connection network DenseNet in a CNN network.
Fig. 3 is a schematic diagram of the inclusion structure of the broadband network in the CNN network.
Fig. 4 is a schematic diagram of a linear CNN structure coding structure based on depth and width modules according to the present application.
FIG. 5 is a schematic diagram of Feature structure in the present application.
Fig. 6 is a schematic diagram of a network structure of a network that is automatically designed by using the network automatic search method FBE-CNN with a noise level of 15 on the Set12 and BSD68 data sets in an embodiment of the present application.
Fig. 7 is a network structure of the network, which is automatically designed by using the network automatic search method FBE-CNN provided in an embodiment of the present application, on the Set12 and BSD68 data sets with a noise level of 25.
Fig. 8 is a network structure of the network, which is automatically designed by using the network automatic search method FBE-CNN provided in an embodiment of the present application, on the Set12 and BSD68 data sets with a noise level of 50.
Detailed Description
To make the objects, technical solutions and advantages of the present invention more apparent, embodiments of the present invention will be described in detail with reference to the accompanying drawings.
The first embodiment is as follows:
the embodiment provides a network automatic searching method aiming at image noise reduction, which comprises the following steps:
Step 2, dividing the public image noise reduction data set according to a preset ratio, and calculating the fitness value of each CNN network individual in the population on the divided data set; training each CNN network individual on a divided data set for 30 epochs to obtain a peak signal-to-noise ratio (PSNR) as an adaptability value corresponding to the CNN network individual;
step 3, performing crossover and mutation operation according to the individual fitness value of each CNN network to generate filial generations;
step 4, generating a new population P by naturally selecting CNN individuals of the offspring and the father;
and 5, repeating the steps 2-4 until a termination condition is met.
Example two
The embodiment provides an image noise reduction method, which searches a network structure for image noise reduction by using the network automatic search method provided in the first embodiment, and the method includes:
Step 2.1: 5, dividing the Gaussian noise reduction data Set12 and the BM3D data Set according to the ratio of 1:4, dividing a test set and a training set according to the ratio, and training 30 epochs on the divided data sets to obtain the peak signal-to-noise ratio (PSNR) on the divided data sets of each CNN individual as the individual fitness;
step 3, performing crossover and mutation operation according to the individual fitness value of each CNN network to generate filial generations;
step 4, generating a new population P by natural selection of CNN individuals of the offspring and the father;
step 5, repeating the steps 2-4 until the termination condition is met
According to the method, depth modules such as ResNet and DenseNet are used, an inclusion width module is introduced, a Feature module is provided as a first layer of a CNN network structure to ensure that an input picture has enough features to learn deep modules, the Feature dimensions of the Feature diagram are transformed by the Transition module, and overfitting of the network is prevented by the Dropout module, wherein the structures of the ResNet module, the DenseNet module and the inclusion module are similar to those of figures 1, 2 and 3, and the Feature module used in the method is provided in figure 5 and is the superposition of Conv + Rlu + BN layers. The present invention performs a linear coding scheme based on the above modules, and connects them into different CNN network individuals, as shown in fig. 4.
In order to prove the effectiveness of the present invention, the present embodiment performs experimental verification on the gaussian noise reduction data Set12 and the BSD68 data Set, and compares the current mainstream gaussian noise reduction methods, for example, including BM3D algorithm and WNNM algorithm based on the nonlinear similarity principle, EPLL algorithm based on the generation model, MLP algorithm based on the independent training principle, CSF algorithm and TNRD algorithm, and DnCNN algorithm.
In the above mentioned current mainstream gaussian noise reduction method, the BM3D algorithm based on the nonlinear similarity principle is introduced as described in "k.dabov, a.foi, v.katkovnik, and k.egiazarin," Image differentiating by space 3-d transform-domain collective filtering, "IEEE Transactions on Image Processing, vol.16, No.8, pp.2080-2095,2007";
the WNNM algorithm can be introduced by reference to the introduction in "S.Gu, L.Zhang, W.Zuo, and X.Feng", "Weighted nuclear norm minimization with application to image differentiation", "in 2014IEEE Conference on Computer Vision and Pattern Recognition,2014, pp.2862-2869";
the introduction of the model-based EPLL algorithm can be referred to as the introduction in "D.Zoran and Y.Weiss," From learning models of natural image schedules to world image retrieval, "in 2011International Conference on Computer Vision,2011, pp.479-486";
reference is made to "h.c. burger, c.j.schuler, and s.harmeling," Image differentiating: Can plant neural networks complete with bm3 d? An introduction in "in 2012IEEE Conference on Computer Vision and Pattern Recognition,2012, pp.2392-2399.";
the CSF algorithm and TNRD algorithm are described in U.S. Schmidt and S.Roth, "reducing fields for effective image restoration," in 2014IEEE Conference on Computer Vision and Pattern Recognition,2014, pp.2774-2781, "and" Y.Chen and T.Pock, "transport non-operation differentiation: A flex frame for fast and effective image restoration," IEEE Transactions on Pattern Analysis and Machine Analysis, vol.39, No.6, pp.6-1272,2017 ";
the DnCNN algorithm is described in "K.Zhang, W.Zuo, Y.Chen, D.Meng, and L.Zhang," Beyond a gaussian noise ": responsive learning of deep cnn for Image noise," IEEE Transactions on Image Processing, vol.26, No.7, pp.3142-3155,2017.
The experimental effect on the gaussian noise reduction dataset Set12 is shown in table 1 below:
TABLE 1 comparison of the Effect of FBE-CNN and the conventional Gaussian noise reduction method on Set12
| BM3D | WNNM | EPLL | MLP | CSF | TNRD | DnCNN | FBE-CNN | |
| θ＝15 | 32.372 | 32.696 | 32.138 | – | 32.318 | 32.502 | 32.839 | 32.49 |
| θ＝25 | 29.969 | 30.257 | 29.692 | 30.027 | 29.837 | 30.055 | 30.378 | 30.28 |
| θ＝50 | 26.722 | 27.052 | 26.471 | 26.783 | – | 26.812 | 27.165 | 27.19 |
In table 1, θ represents a noise level, wherein the larger the value of θ is, the more difficult the noise reduction task is, three noise levels are adopted on the Set12 data Set, which are 15, 25 and 30 respectively, and the corresponding values in each algorithm are percentage values of PSNR average noise ratios corresponding to the noise levels, wherein the higher the PSNR value is, the better the effect of processing the gaussian noise reduction task is.
As can be seen from table 1, the PSNR values of the present invention correspond to 32.9%, 30.28% and 27.19% at three noise levels, respectively.
At a noise level of 15, the effect of the present invention is better than 32.372% of the BM3D method, 32.138% of the EPLL, and 32.502% of the TNRD.
At a noise level of 25, the effect of the present invention is better than 29.969% for BM3D method, 30.297% for WNNM method, 29.629% for EPLL method, 30.027% for MLP, 29.837% for CSF and 30.055% for TNRD.
At the noise level 50, the present invention achieves an effect superior to all algorithms. Although the effect of the present invention is no better than that of DnCNN at noise levels 15 and 25, it is better than that at noise level 50. This shows that, the present invention adopts a structure more complex than the traditional structure based on the linear CNN, because the present application adopts the depth module, ResNet and densneet, and the width inclusion module of the CNN, the present invention can automatically design a more complex structure, and better solve the problem of gaussian noise reduction.
Table 2 gives the behaviour of the invention on the gaussian noise reduction data set BSD 68:
TABLE 2 comparison of the effect of FBE-CNN and the conventional Gaussian noise reduction method on Set12
| BM3D | WNNM | EPLL | MLP | CSF | TNRD | DnCNN | FBE-CNN | |
| θ＝15 | 31.07 | 31.37 | 31.21 | – | 31.24 | 31.42 | 31.73 | 31.56 |
| θ＝25 | 28.57 | 28.83 | 28.68 | 28.96 | 28.74 | 28.92 | 28.68 | 28.95 |
| θ＝50 | 25.62 | 25.87 | 25.67 | 26.03 | – | 25.97 | 26.02 | 26.09 |
As shown in table 2, the PSNR values of the FBE-CNN of the present invention at noise levels of 15, 20 and 50 were 31.56%, 28.95% and 26.09%, respectively.
At noise level 15, the noise reduction effect of the present invention was not as good as that of DnCNN, but 31.07% better than BM3D, 31.37% of WNNM, 31.21% of EPLL, 31.24% of CSF, and 31.42% better than TNRD were achieved.
At noise levels 25 and 50, the noise reduction achieved by the present invention is optimal, respectively.
Similar to the experimental effect of the Set12 data Set, the invention adopts a depth module based on ResNet and DenseNet, an inclusion width module, a Feature extraction module, a Transition Feature conversion module and a Dropout over-fitting prevention module, and has a more complex structure compared with the traditional CNN noise reduction network based on a simple linear module, and can often obtain a very good effect in the face of a difficult Gaussian noise reduction task. Meanwhile, the Gaussian noise reduction task on the BSD68 noise reduction data Set is a task difficult to Set12, so that the best effect can be achieved under the noise levels of 25 and 50, and the superiority of the method for solving the complex Gaussian noise reduction problem is also shown.
Meanwhile, as the automatic structure design algorithm based on the genetic algorithm is adopted, for the Set12 and BSD68 data sets, the invention can automatically adjust the automatically designed network structure according to the problem of different noise levels under the data sets, thereby avoiding the time cost problem of human intervention, the CNN Gaussian network structures automatically designed on the Set12 and BSD68 data sets are as shown in figures 6, 7 and 8, the CNN network structures designed under different noise levels are different, which shows that the invention can adaptively design the optimal CNN network structure for the noise reduction tasks of different noise levels of different data sets, does not need the intervention of human factors, and greatly improves the speed and performance for processing the noise reduction tasks.
Some steps in the embodiments of the present invention may be implemented by software, and the corresponding software program may be stored in a readable storage medium, such as an optical disc or a hard disk.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (8)
1. A method for automatic network search for image noise reduction, the method comprising:
step 1, carrying out variable length linear coding on the CNN structure based on a ResNet module, a DenseNet module, an inclusion module, a Feature module, a Transition module and a Dropout module to obtain a plurality of CNN networks with different structures, and respectively taking the obtained CNN networks with different structures as an initial population P0The first layer of each CNN network individual is a Feature module, and the Transition module is the superposition of Conv + Rlu + BN layers;
step 2, dividing the public image noise reduction data set according to a preset ratio, and calculating the fitness value of each CNN network individual in the population on the divided data set;
step 3, performing crossover and mutation operation according to the fitness value of each CNN network individual to generate filial generations;
step 4, generating a new population P by natural selection of CNN network individuals of the offspring and the father;
and 5, repeating the steps 2-4 until a termination condition is met, wherein the finally obtained CNN network individual is a searched network structure for image noise reduction.
2. The method of claim 1, wherein the step 2 of calculating the fitness value of each CNN network individual in the population on the partitioned data sets comprises:
and (3) training each CNN network individual on a divided data set by 30epoch to obtain a peak signal-to-noise ratio (PSNR) as a fitness value corresponding to the CNN network individual.
3. The method of claim 2, wherein the Feature module comprises two branches, ConvRelu1, ConvRelu2, and ConvRelu 3; ConvRelu3 corresponds to a convolution kernel size of 3 x 3, Padding of 1, move step Stride of 1, Output Channel of c, Output Channel of 32.
4. The method of claim 3 wherein the Transition module takes Conv, BN and ReLU operations where Conv has a convolution kernel size of 1 x 1, Padding Padding of 0, and a move step Stride of 1.
5. The method according to claim 4, wherein the image denoising datasets disclosed in step 2 are Set12 and BM3D datasets.
6. The method of claim 5, wherein the predetermined ratio is 1: 5.
7. The method of claim 6, wherein prior to calculating the fitness value for each CNN network individual in the population on the partitioned data set, further comprises partitioning the test set and the training set on the partitioned data set in a 1:4 ratio.
8. An image noise reduction method, characterized in that the method adopts the network structure for image noise reduction searched by the method of any claim 1-7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210305751.9A CN114663312A (en) | 2022-03-24 | 2022-03-24 | Network automatic searching method aiming at image noise reduction and image noise reduction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210305751.9A CN114663312A (en) | 2022-03-24 | 2022-03-24 | Network automatic searching method aiming at image noise reduction and image noise reduction method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114663312A true CN114663312A (en) | 2022-06-24 |
Family
ID=82031205
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210305751.9A Pending CN114663312A (en) | 2022-03-24 | 2022-03-24 | Network automatic searching method aiming at image noise reduction and image noise reduction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114663312A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117173037A (en) * | 2023-08-03 | 2023-12-05 | 江南大学 | Neural network structure automatic search method for image noise reduction |
-
2022
- 2022-03-24 CN CN202210305751.9A patent/CN114663312A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117173037A (en) * | 2023-08-03 | 2023-12-05 | 江南大学 | Neural network structure automatic search method for image noise reduction |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109711426B (en) | Pathological image classification device and method based on GAN and transfer learning | |
| CN113011570B (en) | Facial expression recognition method adopting neural network compression system | |
| CN110175628A (en) | A kind of compression algorithm based on automatic search with the neural networks pruning of knowledge distillation | |
| CN107392919B (en) | Adaptive genetic algorithm-based gray threshold acquisition method and image segmentation method | |
| CN112883839B (en) | Remote sensing image interpretation method based on adaptive sample set construction and deep learning | |
| CN112465120A (en) | Fast attention neural network architecture searching method based on evolution method | |
| CN111008224B (en) | Time sequence classification and retrieval method based on deep multitasking representation learning | |
| CN111488917A (en) | Garbage image fine-grained classification method based on incremental learning | |
| CN111353534B (en) | Graph data category prediction method based on adaptive fractional order gradient | |
| CN110516724B (en) | High-performance multi-layer dictionary learning characteristic image processing method for visual battle scene | |
| CN111988329A (en) | Network intrusion detection method based on deep learning | |
| CN113032613B (en) | Three-dimensional model retrieval method based on interactive attention convolution neural network | |
| Liu et al. | EACP: An effective automatic channel pruning for neural networks | |
| CN114118369B (en) | Image classification convolutional neural network design method based on group intelligent optimization | |
| Soudy et al. | RepConv: A novel architecture for image scene classification on Intel scenes dataset | |
| Lata et al. | Data augmentation using generative adversarial network | |
| CN114663312A (en) | Network automatic searching method aiming at image noise reduction and image noise reduction method | |
| CN117290721A (en) | Digital twin modeling method, device, equipment and medium | |
| CN114972794A (en) | Three-dimensional object recognition method based on multi-view Pooll transducer | |
| CN118628736A (en) | Weak supervision indoor point cloud semantic segmentation method, device and medium based on clustering thought | |
| CN114140657A (en) | Image retrieval method based on multi-feature fusion | |
| CN117349687A (en) | Daily load curve clustering method based on variable convolution self-encoder | |
| CN114581470B (en) | Image edge detection method based on plant community behaviors | |
| CN115908697A (en) | Generation model based on point cloud probability distribution learning and method thereof | |
| CN115063374A (en) | Model training method, face image quality scoring method, electronic device and storage medium |
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 |