CN110648292A - High-noise image denoising method based on deep convolutional network - Google Patents
High-noise image denoising method based on deep convolutional network Download PDFInfo
- Publication number
- CN110648292A CN110648292A CN201910860560.7A CN201910860560A CN110648292A CN 110648292 A CN110648292 A CN 110648292A CN 201910860560 A CN201910860560 A CN 201910860560A CN 110648292 A CN110648292 A CN 110648292A
- Authority
- CN
- China
- Prior art keywords
- image
- noise
- network
- size
- convolution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 45
- 230000006870 function Effects 0.000 claims abstract description 25
- 230000004913 activation Effects 0.000 claims abstract description 14
- 238000009826 distribution Methods 0.000 claims abstract description 8
- 238000012549 training Methods 0.000 claims description 14
- 238000010606 normalization Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 239000000654 additive Substances 0.000 claims description 5
- 230000000996 additive effect Effects 0.000 claims description 5
- 238000005457 optimization Methods 0.000 claims description 5
- 230000008901 benefit Effects 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 210000002569 neuron Anatomy 0.000 claims description 2
- 230000003190 augmentative effect Effects 0.000 claims 1
- 230000003247 decreasing effect Effects 0.000 abstract description 2
- 238000000926 separation method Methods 0.000 abstract description 2
- 230000000007 visual effect Effects 0.000 abstract 1
- 230000000694 effects Effects 0.000 description 7
- 238000013135 deep learning Methods 0.000 description 5
- 238000000605 extraction Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000009792 diffusion process Methods 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000012360 testing method Methods 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
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention relates to a high-noise image denoising method based on a deep convolutional network. The method comprises the steps of firstly, extracting the characteristics of an image containing noise by adopting incremental expansion convolution, batch standardization operation and a Leakly ReLU function; then, restoring the image by adopting a mode of combining a decreasing expansion volume and a ReLU activation function; then, the separation of the network model from the image noise and the content is realized through the combination of residual learning and batch standardization operation; finally, learning the optimal weight parameter of the network model by solving the value of the minimized loss function (adopting different loss functions aiming at different noise distributions); and finally denoising the noise image by using the trained network model. The invention can effectively remove the image noise in the high-noise environment, improve the visual effect of the image and has better practicability.
Description
Technical Field
The invention relates to the field of image processing and computer vision, mainly relates to an image denoising method in the field of deep learning, and particularly relates to a high-noise image denoising method based on a deep convolutional network.
Background
With deep learning, particularly the successful application of the convolutional neural network in the fields of image feature extraction, identification and the like, a new direction is provided for the image denoising problem, particularly the image denoising problem in a high-noise environment. Compared with the traditional classical image denoising method, the denoising method based on the deep convolutional network has stronger learning capacity, can effectively improve the adaptability of the network model to different noise distributions and different noise levels by training through a large amount of noise image sample data, and has better denoising generalization capacity. In 2012, Xie et al applied the stacked sparse denoising self-encoder method to solve the gaussian noise elimination and realized the denoising effect equivalent to K-SVD (a dictionary learning algorithm); in 2016, Chen et al proposed a trainable, nonlinear reaction diffusion model, TNRD, that improves the de-noising performance of images by unfolding a fixed number of gradient descent feedforward depth networks; in 2017, Zhang et al propose a deep learning based denoising model DnCNN, the method adopts a single denoising model for training to realize the denoising task of images, meanwhile, the method has a good denoising effect on images with unknown noise levels, and the network model proves that the denoising performance and efficiency of the network model are superior to those of a classical denoising method BM3D algorithm through experiments.
The image denoising method has different choices of the training set, the training target and the feature, and can obtain good denoising effect in a low-noise environment. However, the denoising effect of these methods in a high noise environment is not ideal. In order to further optimize the image denoising effect in a high-noise environment, an image denoising method of a symmetric extended convolution residual error network is provided.
Disclosure of Invention
The invention provides a symmetrical extended convolution residual network image denoising method, which comprises the steps of firstly selecting and extracting characteristics of an input noise image through a convolution network with a symmetrical structure, then reconstructing the extracted image characteristics, finally realizing effective separation of noise and image content through the combination of residual learning and batch standardization, and outputting a residual image with the same size as an original image. The method effectively solves the problem of covariate transfer inside the network by using batch standardization operation, reduces the problem of image boundary artifacts by adopting the operation of zero filling on the non-convolved image, and obviously improves the noise reduction performance and the imaging quality.
The invention adopts the following technical scheme:
a high-noise image denoising method based on a deep convolutional network specifically comprises the following steps:
step S1: selecting a data set;
step S2: preprocessing the selected data set;
step S3: establishing a symmetrical extended convolution residual error network by combining the noise type in the image;
step S4: sending the noisy image and the clear label image corresponding to the noisy image into the symmetric extended convolution residual error network to obtain an image denoising network model;
step S5: and (4) learning the optimal parameters of the network model by solving the value of the minimum loss function, and restoring the noise image by using the trained network model.
Further, step S2 specifically includes the following steps:
step S21: taking the noiseless images with the size of 180 x 180 in the data set as an original training set, and simultaneously performing data enhancement on the noiseless images to obtain 128 x 1600 sample images with the size of 53 x 53;
step S22: creating different noise models, and adding different noises to the data set according to requirements;
further, step S22 specifically includes the following steps:
step S221: designing three noise models which are additive Gaussian noise respectively, wherein the standard deviation sigma of model parameter noise belongs to [0,60 ]; poisson noise, wherein the model parameter noise amplitude lambda belongs to [0,60 ]; the method comprises the following steps of (1) carrying out multiplicative Bernoulli noise, wherein p represents the damage probability of a pixel, and then setting model parameters p belonging to [0,0.95 ];
step S222: selecting an image from the data set and sending the image into a noise model, namely adding noise distribution;
step S223: iterating step S222 to obtain a group of noisy training sets;
further, step S3 specifically includes the following steps:
step S31: the network structure makes a trade-off between the size of the receptive field and the depth of the network using a symmetric extended convolution residual network. The mode from the first layer to the second last layer is the expanding convolution, the expanding factors (the expanding factors are formed by sparse filters with the size of (2r +1) × (2r +1) (r represents the depth of the network)) are respectively set to be 1, 2, 3, 4, 5, 4, 3, 2, 1, and the mode of increasing the area of the receptive field to obtain more context information because damaged pixels in the image are recovered is a widely used method, and the expanding convolution not only has the capability of expanding the receptive field, but also has the advantage of the classical 3 × 3 convolution;
step S32: the last layer of the network model uses a convolution with a convolution kernel size of 3 x 3. Specifically, the convolution layer realizes the reconstruction of the characteristic image and outputs the characteristic image through c (image channel number) convolution kernels of 3 × 32 and the step length of 1 × 1;
further, step S31 specifically includes the following steps:
step S311: at the first level in the network, the used expansion rate size is 1, the convolution kernel size is 1d × c, the number of convolution kernels is 32, the step size is set to 1, and the used activation function is leak lu (max (0.01x, x)). Wherein d represents an expansion factor, c represents the number of channels of the image, and x represents the input characteristic;
step S312: in the second to fifth layers of the network, using expansion rates of size 2, 3, 4, 5 respectively, using 64 convolution kernels of size 2d x 32, 128 convolution kernels of size 3d x 64, 256 convolution kernels of size 4d x 128 and c convolution kernels of size 5d x 256, respectively, with a step size set to 1, using an activation function of leak ReLU (max (0.01x, x)), using a batch normalization method between the expanded convolution layer and the activation function;
step S313: in the sixth to ninth layers of the network, the used expansion rates are respectively 4, 3, 2 and 1, 256 convolution kernels with the size of 4d × c, 128 convolution kernels with the size of 3d × 256, 64 convolution kernels with the size of 2d × 128 and 32 convolution kernels with the size of 1d × 64 are respectively used, the step size is set to be 1, and the used activation function is a ReLU activation function;
further, step S312 specifically includes the following steps:
step S3121: the batch normalization calculation formula is as follows:
wherein gamma and beta represent adjustable parameters,
is K after gamma and beta parametrizationnormDistribution of (2)
Step S3122: knormIs defined as follows:
wherein, KnormRepresenting the regularization results, unactivated neuron nodes, μ and σ, in the K-network2Representing the mean and variance of the samples, respectively, and ξ represents a very small, non-zero positive number in order to ensure that the denominator is meaningful. The calculation process is used for the image characteristics extracted by each convolution layer, and normalization processing is carried out on the characteristic data, so that the same characteristic data distribution can be learned by each layer of the network, and the network convergence speed and the training efficiency are improved.
Further, step S5 specifically includes the following steps:
step S51: different image noises adopt different Loss functions, and Loss adopted by additive Gaussian noise:
poisson noise exploits losses with L2:
the loss employed by the bernoulli noise is:
wherein f represents the denoised image, X represents a clear label image, n represents the number of samples of the training batch, w and h represent the width and height of the image, and yiRepresentation entityThe value of the actual value is the value,
a random value representing the predicted output value of the network, v being 0 or 1;
step S52: the residual map is trained according to the process of residual learning, and then the MSE (mean square error) between the desired residual image and the predicted residual image is calculated:
wherein, P represents a clean label image, Q represents a noise image, N represents the number of sample pairs, R represents a noise image predicted by a network model, and w and b represent weight parameters and bias terms of the network respectively. Obtaining errors of each layer according to a back propagation process, and adjusting weight parameters of each layer according to the errors to complete optimization of the network model;
step S53: the steps S51 and S52 are iterated until the network converges.
Compared with the prior art, the invention has the following advantages: the invention provides a high-noise image denoising method based on a deep convolutional network, which can better finish the image denoising problem under a high-noise environment by constructing a symmetric extended convolutional residual error network and utilizing a Leakly ReLU and ReLU dual activation function, and meanwhile, the obtained restored image does not have the problems of loss of edge information and boundary artifacts.
Drawings
FIG. 1 is a schematic flow chart of a method according to an embodiment of the present invention.
Fig. 2 is a structure diagram of a class-one feature extraction layer according to an embodiment of the present invention.
Fig. 3 is a structure diagram of a class two feature extraction layer according to an embodiment of the present invention.
Fig. 4 is a diagram of a network model structure for image denoising according to an embodiment of the present invention.
Detailed Description
As shown in fig. 1, the present embodiment provides a method for denoising a high-noise image based on a deep convolutional network, which includes:
step S1: selecting a data set;
step S2: preprocessing the selected data set;
step S3: establishing a symmetrical extended convolution residual error network by combining the noise type in the image;
step S4: sending the noisy image and the clear label image corresponding to the noisy image into the symmetric extended convolution residual error network to obtain an image denoising network model;
step S5: and (4) learning the optimal parameters of the network model by solving the value of the minimum loss function, and restoring the noise image by using the trained network model.
In step S1, the present embodiment uses 300 images in the BSDS300 as a training Set, and uses Set68 as a test Set, which includes 68 natural images.
In step S2, the present embodiment uses additive gaussian noise, where the model parameter noise standard deviation σ ∈ [0,60 ]; poisson noise, wherein the model parameter noise amplitude lambda belongs to [0,60 ]; and the Bernoulli noise is represented by p, and the damage probability of the pixel is represented by p, the model parameter p belongs to 0, 0.95.
In steps S3, S4, S5, the structure of the deep convolutional neural network used is as shown in fig. 4. The method comprises the steps that a noise image is subjected to feature learning of an image through an incremental expansion convolution part of a network model, meanwhile, the problem of internal covariate transfer of the network after convolution is solved by batch standardization operation after each layer of convolution, and image feature nonlinear mapping is carried out through a Leaky ReLU activation function to realize selection and extraction of image features; then, reconstructing image characteristics of the obtained characteristic image through a decreasing type expansion convolution part of a network model, batch standardization operation and nonlinear mapping of a ReLU activation function, and outputting a residual image; then, calculating the obtained residual image through a specified loss function; and finally, carrying out reverse iteration updating and optimization on the corresponding network weight parameters by using an optimization algorithm to obtain the optimal parameters of the network.
In step S5, a loss function is optimized in a training phase by using an SGD algorithm (Stochastic Gradient Descent) algorithm, and a network optimal parameter is obtained through the optimization of the loss function, and finally, an input noise image is denoised by using a network model.
The invention provides a high-noise image denoising method based on a deep convolutional network with better performance by using a symmetric extended convolutional residual network model and combining the idea of residual learning through the thinking of the traditional image denoising method and the current image denoising method combined with the deep learning, namely, the method has better denoising effect in a low-noise environment, but the denoising effect in a high noise level is not ideal, and the problem of boundary artifacts is faced when the deep learning is used for denoising images.
The above description is a better embodiment of the present invention, and is not intended to limit the present invention in other forms. It will be appreciated by those skilled in the art that the foregoing embodiments and descriptions have been presented only to illustrate the principles and advantages of the invention, and that various modifications may be made without departing from the spirit and scope of the invention, which changes and modifications are intended to be within the scope of the invention as claimed.
Claims (7)
1. A high-noise image denoising method based on a deep convolutional network is characterized by comprising the following steps: comprises the following steps:
step S1: selecting a data set;
step S2: preprocessing the selected data set;
step S3: establishing a symmetrical extended convolution residual error network by combining the noise type in the image;
step S4: sending the noisy image and the corresponding clear label image into the symmetric extended convolution residual error network to obtain an image denoising network model;
step S5: and (4) learning the optimal parameters of the network model by solving the value of the minimum loss function, and restoring the noise image by using the trained network model.
2. The method for denoising high-noise image based on deep convolutional network as claimed in claim 1, wherein: step S2 specifically includes the following steps:
step S21: taking the noiseless images with the size of 180 x 180 in the data set as an original training set, and simultaneously performing data enhancement on the noiseless images to obtain 128 x 1600 sample images with the size of 53 x 53;
step S22: different noise models are created, and different noises are added to the data set according to requirements.
3. The method for denoising a high-noise image based on a deep convolutional network as claimed in claim 2, wherein: step S22 specifically includes the following steps:
step S221: designing three noise models which are additive Gaussian noise respectively, wherein the standard deviation sigma of model parameter noise belongs to [0,60 ]; poisson noise, wherein the model parameter noise amplitude lambda belongs to [0,60 ]; the method comprises the following steps of (1) carrying out multiplicative Bernoulli noise, wherein p represents the damage probability of a pixel, and then setting model parameters p belonging to [0,0.95 ];
step S222: selecting an image from the data set and sending the image into a noise model, namely adding noise distribution;
step S223: step S222 is iterated to obtain a noisy training set.
4. The method for denoising high-noise image based on deep convolutional network as claimed in claim 1, wherein: step S3 specifically includes the following steps:
step S31: a symmetric augmented convolutional residual network is used. The network structure is balanced between the size of the receptive field and the depth of the network, and the modes from the first layer to the second last layer are extended convolution, and the extension factors (the extension factors are formed by sparse filters with the size of (2r +1) × (2r +1) (r represents the depth of the network)) are respectively set to be 1, 2, 3, 4, 5, 4, 3, 2 and 1. Because damaged pixels in the image need to be recovered, a mode of increasing the area of the receptive field to obtain more context information is a widely used method, and the expansion convolution not only has the capability of expanding the receptive field, but also has the advantage of classical 3-by-3 convolution;
step S32: the last layer of the network model uses a convolution with a convolution kernel size of 3 x 3. Specifically, the convolution layer reconstructs and outputs a feature image by using c (number of image channels) convolution kernels of 3 × 32 and a step size of 1 × 1.
5. The method of claim 4, wherein the method comprises: step S31 specifically includes the following steps:
step S311: at the first level in the network, the used expansion rate size is 1, the convolution kernel size is 1d × c, the number of convolution kernels is 32, the step size is set to 1, and the used activation function is leak lu (max (0.01x, x)). Wherein d represents an expansion factor, c represents the number of channels of the image, and x represents the input characteristic;
step S312: in the second to fifth layers of the network, using expansion rates of size 2, 3, 4, 5 respectively, using 64 convolution kernels of size 2d x 32, 128 convolution kernels of size 3d x 64, 256 convolution kernels of size 4d x 128 and c convolution kernels of size 5d x 256, respectively, with a step size set to 1, using an activation function of leak ReLU (max (0.01x, x)), using a batch normalization method between the expanded convolution layer and the activation function;
step S313: in the sixth to ninth layers of the network, the expansion rates used were 4, 3, 2, 1, respectively, 256 convolution kernels of size 4d × c, 128 convolution kernels of size 3d × 256, 64 convolution kernels of size 2d × 128, 32 convolution kernels of size 1d × 64, the step size was set to 1, and the activation function used was the ReLU activation function.
6. The method of claim 5, wherein the method comprises: step S312 specifically includes the following steps:
step S3121: the batch normalization calculation formula is as follows:
wherein gamma and beta represent adjustable parameters,
is K after gamma and beta parametrizationnormThe distribution of (a);
step S3122: knormIs defined as follows:
wherein, KnormRepresenting the regularization result, K represents the unactivated neuron nodes in the network, μ and σ2Representing the mean and variance of the samples, respectively, and ξ represents a very small, non-zero positive number in order to ensure that the denominator is meaningful. The calculation process is used for the image characteristics extracted by each convolution layer, and normalization processing is carried out on the characteristic data, so that the same characteristic data distribution can be learned by each layer of the network, and the convergence speed and the training efficiency are improved.
7. The method for denoising high-noise image based on deep convolutional network as claimed in claim 1, wherein: step S5 specifically includes the steps of:
step S51: different image noise uses different Loss functions. Loss adopted by additive gaussian noise:
poisson noise exploits losses with L2:
the loss employed by the bernoulli noise is:
wherein f represents the denoised image, X represents a clear label image, n represents the number of samples of the training batch, w and h represent the width and height of the image, and yiThe actual value is represented by a value that is,
a random value representing the predicted output value of the network, v being 0 or 1;
step S52: the residual map is trained according to the process of residual learning, and then the MSE (mean square error) between the desired residual image and the predicted residual image is calculated:
wherein, P represents a clean label image, Q represents a noise image, N represents the number of sample pairs, R represents a noise image predicted by a network model, and w and b represent weight parameters and bias terms of the network respectively. Obtaining errors of each layer according to a back propagation process, and adjusting weight parameters of each layer according to the errors to complete optimization of the network model;
step S53: the steps S51 and S52 are iterated until the network converges.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910860560.7A CN110648292B (en) | 2019-09-11 | 2019-09-11 | High-noise image denoising method based on deep convolutional network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910860560.7A CN110648292B (en) | 2019-09-11 | 2019-09-11 | High-noise image denoising method based on deep convolutional network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110648292A true CN110648292A (en) | 2020-01-03 |
| CN110648292B CN110648292B (en) | 2022-06-21 |
Family
ID=68991204
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910860560.7A Active CN110648292B (en) | 2019-09-11 | 2019-09-11 | High-noise image denoising method based on deep convolutional network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110648292B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111612709A (en) * | 2020-05-11 | 2020-09-01 | 济南浪潮高新科技投资发展有限公司 | Image noise reduction method based on DnCNNs improvement |
| CN111899188A (en) * | 2020-07-08 | 2020-11-06 | 西北工业大学 | Neural network learning cone beam CT noise estimation and suppression method |
| CN111915513A (en) * | 2020-07-10 | 2020-11-10 | 河海大学 | Image denoising method based on improved adaptive neural network |
| CN112580554A (en) * | 2020-12-25 | 2021-03-30 | 北京环境特性研究所 | CNN-based MSTAR data noise intensity control classification identification method |
| CN114299004A (en) * | 2021-12-25 | 2022-04-08 | 宁波凯德科技服务有限公司 | Image semantic segmentation method based on pepper salt noise total variation data augmentation |
| CN115880187A (en) * | 2023-01-10 | 2023-03-31 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Single-image reflection removing method based on denoising diffusion probability model and related equipment |
| CN116993845A (en) * | 2023-06-09 | 2023-11-03 | 西安交通大学 | CT image artifact removal method based on integrated depth network DnCNN |
| CN117095074A (en) * | 2023-08-23 | 2023-11-21 | 浙江大学 | Medical image double denoising method and device based on deep learning |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070177817A1 (en) * | 2006-01-27 | 2007-08-02 | Microsoft Corporation | Region-based image denoising |
| CN107689034A (en) * | 2017-08-16 | 2018-02-13 | 清华-伯克利深圳学院筹备办公室 | A kind of training method of neutral net, denoising method and device |
| CN108416755A (en) * | 2018-03-20 | 2018-08-17 | 南昌航空大学 | A kind of image de-noising method and system based on deep learning |
| CN108830809A (en) * | 2018-06-05 | 2018-11-16 | 陕西师范大学 | One kind is based on expansion convolved image denoising method |
| CN109636734A (en) * | 2018-10-29 | 2019-04-16 | 昆明理工大学 | A kind of sparse regular terms of weighting based on group rarefaction representation constrains image de-noising method |
| CN109658344A (en) * | 2018-11-12 | 2019-04-19 | 哈尔滨工业大学（深圳） | Image de-noising method, device, equipment and storage medium based on deep learning |
| CN110163827A (en) * | 2019-05-28 | 2019-08-23 | 腾讯科技（深圳）有限公司 | Training method, image de-noising method, device and the medium of image denoising model |
-
2019
- 2019-09-11 CN CN201910860560.7A patent/CN110648292B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20070177817A1 (en) * | 2006-01-27 | 2007-08-02 | Microsoft Corporation | Region-based image denoising |
| CN107689034A (en) * | 2017-08-16 | 2018-02-13 | 清华-伯克利深圳学院筹备办公室 | A kind of training method of neutral net, denoising method and device |
| CN108416755A (en) * | 2018-03-20 | 2018-08-17 | 南昌航空大学 | A kind of image de-noising method and system based on deep learning |
| CN108830809A (en) * | 2018-06-05 | 2018-11-16 | 陕西师范大学 | One kind is based on expansion convolved image denoising method |
| CN109636734A (en) * | 2018-10-29 | 2019-04-16 | 昆明理工大学 | A kind of sparse regular terms of weighting based on group rarefaction representation constrains image de-noising method |
| CN109658344A (en) * | 2018-11-12 | 2019-04-19 | 哈尔滨工业大学（深圳） | Image de-noising method, device, equipment and storage medium based on deep learning |
| CN110163827A (en) * | 2019-05-28 | 2019-08-23 | 腾讯科技（深圳）有限公司 | Training method, image de-noising method, device and the medium of image denoising model |
Non-Patent Citations (3)
| Title |
|---|
| ASAD MAHMOOD等: "image denoising and residual", 《2018 9TH WORKSHOP ON HYPERSPECTRAL IMAGE AND SIGNAL PROCESSING: EVOLUTION IN REMOTE SENSING (WHISPERS)》 * |
| 吴从中等: "结合深度残差学习和感知损失的图像去噪", 《中国图象图形学报》 * |
| 盖杉等: "基于深度学习的高噪声图像去噪算法", 《自动化学报》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111612709A (en) * | 2020-05-11 | 2020-09-01 | 济南浪潮高新科技投资发展有限公司 | Image noise reduction method based on DnCNNs improvement |
| CN111612709B (en) * | 2020-05-11 | 2023-03-28 | 山东浪潮科学研究院有限公司 | Image noise reduction method based on DnCNNs improvement |
| CN111899188B (en) * | 2020-07-08 | 2022-06-07 | 西北工业大学 | Neural network learning cone beam CT noise estimation and suppression method |
| CN111899188A (en) * | 2020-07-08 | 2020-11-06 | 西北工业大学 | Neural network learning cone beam CT noise estimation and suppression method |
| CN111915513A (en) * | 2020-07-10 | 2020-11-10 | 河海大学 | Image denoising method based on improved adaptive neural network |
| CN112580554A (en) * | 2020-12-25 | 2021-03-30 | 北京环境特性研究所 | CNN-based MSTAR data noise intensity control classification identification method |
| CN112580554B (en) * | 2020-12-25 | 2024-01-05 | 北京环境特性研究所 | Classification recognition method for MSTAR data noise intensity control based on CNN |
| CN114299004A (en) * | 2021-12-25 | 2022-04-08 | 宁波凯德科技服务有限公司 | Image semantic segmentation method based on pepper salt noise total variation data augmentation |
| CN115880187A (en) * | 2023-01-10 | 2023-03-31 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Single-image reflection removing method based on denoising diffusion probability model and related equipment |
| CN116993845A (en) * | 2023-06-09 | 2023-11-03 | 西安交通大学 | CT image artifact removal method based on integrated depth network DnCNN |
| CN116993845B (en) * | 2023-06-09 | 2024-03-15 | 西安交通大学 | CT image artifact removal method based on integrated depth network DnCNN |
| CN117095074A (en) * | 2023-08-23 | 2023-11-21 | 浙江大学 | Medical image double denoising method and device based on deep learning |
| CN117095074B (en) * | 2023-08-23 | 2024-04-09 | 浙江大学 | Medical image double denoising method and device based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110648292B (en) | 2022-06-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110648292B (en) | High-noise image denoising method based on deep convolutional network | |
| CN107633486B (en) | Structural magnetic resonance image denoising method based on three-dimensional full-convolution neural network | |
| CN110232394B (en) | Multi-scale image semantic segmentation method | |
| CN109859147B (en) | Real image denoising method based on generation of antagonistic network noise modeling | |
| CN109101975B (en) | Image semantic segmentation method based on full convolution neural network | |
| CN106991646B (en) | Image super-resolution method based on dense connection network | |
| CN108596258B (en) | Image classification method based on convolutional neural network random pooling | |
| CN109035142B (en) | Satellite image super-resolution method combining countermeasure network with aerial image prior | |
| CN109035149B (en) | License plate image motion blur removing method based on deep learning | |
| CN108985317B (en) | Image classification method based on separable convolution and attention mechanism | |
| CN111369487B (en) | Hyperspectral and multispectral image fusion method, system and medium | |
| CN111275643B (en) | Real noise blind denoising network system and method based on channel and space attention | |
| CN110706181B (en) | Image denoising method and system based on multi-scale expansion convolution residual error network | |
| CN106952228A (en) | The super resolution ratio reconstruction method of single image based on the non local self-similarity of image | |
| CN111028163A (en) | Convolution neural network-based combined image denoising and weak light enhancement method | |
| CN111192211B (en) | Multi-noise type blind denoising method based on single deep neural network | |
| US20210312591A1 (en) | Systems and method of training networks for real-world super resolution with unknown degradations | |
| CN109801218B (en) | Multispectral remote sensing image Pan-sharpening method based on multilayer coupling convolutional neural network | |
| CN111738954B (en) | Single-frame turbulence degradation image distortion removal method based on double-layer cavity U-Net model | |
| CN111861906A (en) | Pavement crack image virtual augmentation model establishment and image virtual augmentation method | |
| CN112435162B (en) | Terahertz image super-resolution reconstruction method based on complex domain neural network | |
| CN111402138A (en) | Image super-resolution reconstruction method of supervised convolutional neural network based on multi-scale feature extraction fusion | |
| CN113538234A (en) | Remote sensing image super-resolution reconstruction method based on lightweight generation model | |
| CN110838089A (en) | Fast image denoising method based on OctBlock dense block | |
| CN113643303A (en) | Three-dimensional image segmentation method based on two-way attention coding and decoding network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |