CN112669210A - Image super-resolution method, device and storage medium based on VDSR model applying novel ReLU function - Google Patents
Image super-resolution method, device and storage medium based on VDSR model applying novel ReLU function Download PDFInfo
- Publication number
- CN112669210A CN112669210A CN202011576889.XA CN202011576889A CN112669210A CN 112669210 A CN112669210 A CN 112669210A CN 202011576889 A CN202011576889 A CN 202011576889A CN 112669210 A CN112669210 A CN 112669210A
- Authority
- CN
- China
- Prior art keywords
- relu function
- image
- improved
- vdsr
- vdsr model
- 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
Images
Abstract
The invention relates to an image super-resolution method, equipment and a storage medium based on a VDSR model applying a novel ReLU function, which are characterized in that: inputting an image to be processed into a trained improved VDSR model, and outputting to obtain a high-resolution image of the image; a novel ReLU function in the improved VDSR model is a ReLU function of a self-adaptive learning static working point; the invention is inspired by a triode amplifying circuit, introduces the concept of a static working point into a novel ReLU function, takes a zero point in the traditional ReLU function as the static working point, and realizes the self-adaptive learning of the value of the static working point in the neural network training process. The novel ReLU function is applied to a VDSR model, and a data augmentation and learning rate attenuation strategy is adopted in the network training process to avoid the overfitting phenomenon of the network. The invention can effectively improve the performance of the VDSR model in the super-resolution task.
Description
Technical Field
The invention relates to an image super-resolution method, equipment and a storage medium based on a VDSR model applying a novel ReLU function, and belongs to the technical field of deep learning.
Background
Deep learning is a collection of high-complexity data modeling algorithms through multilayer nonlinear transformation, and a deep neural network becomes one of the most important research directions in the field of deep learning by virtue of strong learning and expression capabilities of the deep neural network, and is widely applied to the fields of image processing, video processing and the like.
Each neuron node in the neural network receives the output value of the neuron in the previous layer as the input value of the neuron, and transmits the input value to the next layer, and the neuron node in the input layer can directly transmit the input attribute value to the next layer (hidden layer or output layer). In a multi-layer neural network, there is a functional relationship between the output of an upper node and the input of a lower node, and this function is called an activation function (also called an excitation function). Early neural network models did not introduce activation functions, such as Multilayer Perceptron (MPL), and each layer of neurons simply performed linear transformation on the upper layer of neuron outputs, which resulted in no difference in the expression ability between the fully-connected neural network of any layer and the single-layer neural network model. When the activation function is introduced, the neural network model has nonlinear fitting capability, the expression capability of the neural network is greatly enhanced, and almost any function can be fitted.
Because the activation function brings great improvement to the performance of the neural network, researchers have conducted a lot of research on the design of the activation function, from the early Sigmoid function and Tanh function to the recently widely used ReLU function, and the following describes in detail several commonly used activation functions.
First, Sigmoid function, the mathematical form of Sigmoid function is as follows:
the Sigmoid function image is shown in fig. 1; the horizontal and vertical coordinates represent the input and output of the Sigmoid function respectively, the Sigmoid function maps the input to the interval of (0,1), the gradient disappears easily when the gradient propagates in the reverse direction in the deep neural network, the convergence is slow when the model is trained because the output of the Sigmoid function is not the 0 mean value (zero-center), and the operation consumption is large because the Sigmoid function contains power operation.
Second, the Tanh function, the mathematical form of the Tanh function, is as follows:
as shown in FIG. 2, the horizontal and vertical coordinates of the Tanh function image respectively represent the input and output of the Tanh function, and the Tanh function solves the problem that the output of the Sigmoid function is not 0-mean, but the problems of gradient extinction and power operation still exist.
Third, the ReLU function, the mathematical form of the ReLU function, is as follows:
the ReLU function image is shown in fig. 3: the abscissa and the ordinate respectively represent the input and the output of the ReLU function, which is actually a maximum function. Although simple, is an important outcome in recent years. Compared with a Sigmoid function and a Tanh function, the ReLU function does not contain power operation any more, so that the operation consumption is greatly reduced, and the gradient disappearance problem is effectively relieved by the ReLU function. However, the ReLU function still has the problem of outputting a non-zero mean value, and in the case of improper parameter initialization or excessively high initial learning rate, it may result in some neurons in the network never being activated. Despite these problems, the ReLU function is still one of the most popular activation functions at present, and variations such as Leaky ReLU, RReLU, and prilu have appeared in subsequent studies.
In recent years, convolutional neural networks have made great progress in the task of image super-resolution, and researchers often increase the performance by deepening the depth of a network model, for example, a VDSR model includes 20 convolutional layers. If the Sigmoid function or tanh function mentioned in the foregoing is applied as the activation function layer, the training of the network model becomes unstable, and the gradient disappearance easily occurs, so that the training cannot be continued. The original VDSR model selects a ReLU function as an activation function layer, and obtains relatively ideal performance. However, the ReLU function zeroes all the feature data smaller than 0, which limits the expressive power of the model.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides an image super-resolution method based on a VDSR model applying a novel ReLU function;
the invention also provides computer equipment and a computer storage medium.
Interpretation of terms:
1. the VDSR model is a classical image super-resolution network model, and the network structure thereof is shown in fig. 5. Including 20 convolutional layers, 19 ReLU activation function layers, and residual concatenation. The model takes a low-resolution image as input after bicubic interpolation, learns the lost high-frequency component of the low-resolution image through 20 layers of convolution layers, and performs pixel-level addition on the learned high-frequency component and the input at the end of the network to obtain a final high-resolution image.
2. The public data set BSD300, provided by Berkely computer Vision Group, contains 200 training charts and 100 test charts.
The technical scheme of the invention is as follows:
an image super-resolution method based on an improved VDSR model applying a novel ReLU function is as follows: inputting an image to be processed into a trained improved VDSR model, and outputting to obtain a high-resolution image of the image;
the new ReLU function in the improved VDSR model, i.e., the ReLU function f (x) of the adaptive learning static operating point, is shown in formula (i):
in the formula (I), Q refers to a static operating point, the static operating point refers to a zero point in an original ReLU function, Q is obtained through adaptive learning, and x represents the input characteristic of a new ReLU function layer in an improved VDSR model.
Q is obtained through adaptive learning, and means that: setting Q as a learnable parameter and initializing to 0, wherein x represents the input characteristic of a new ReLU function layer in the improved VDSR model, in the training process of the improved VDSR model, the value smaller than Q in the input characteristic is set to 0 by the new ReLU function layer, the value larger than Q is set to x-Q, the Q value is continuously updated in the back propagation process, and after training of a large amount of data, the new ReLU function layers at different positions in the VDSR model have different Q values. Compared with the original ReLU function, the novel ReLU function provided by the invention has better nonlinear fitting capability, and the original ReLU function refers to the ReLU function mentioned in the background art.
The concept of a quiescent operating point stems from the triode amplification principle. The triode has two states, namely a static state and a dynamic state, wherein the static state refers to a direct current working state when no signal is added to the triode, the voltage of each electrode in the static state is called static working current, the dynamic state refers to a working state when an alternating current signal is added to the triode, and the current of each electrode in the dynamic state is called dynamic working current. If the direct current circuit of the triode does not work normally, the alternating current circuit of the triode cannot work normally. The static operating point is that when in a static state, the circuit is in a direct current operating state, and the numerical values of the current and the voltage can be represented by a determined point on an input-output curve of the triode, which is also called as a Q point. The static values of the voltage and the current of the amplifying circuit are determined, and a proper static working point is selected to prevent the circuit from generating nonlinear distortion and ensure the amplifying effect. In the original ReLU function, data with input greater than 0 is output as the original ReLU function, and the input characteristic less than 0 sets the output to 0, and in combination with the concept of the quiescent operating point in the triode, the zero point in the original ReLU function is called the quiescent operating point of the activation function.
Preferably according to the present invention, the improved VDSR model comprises 20 convolutional layers, 19 novel ReLU function layers, and residual concatenation; the improved VDSR model takes the low-resolution image as input after bicubic interpolation, learns the lost high-frequency component of the low-resolution image through 20 layers of convolution layers, and performs pixel-level addition on the learned high-frequency component and the input at the end of the improved VDSR model to obtain the final high-resolution image.
The improved VDSR model is identical in network structure to the original VDSR model, except that the original ReLU function is replaced with the new ReLU function proposed by the present invention.
Preferably, according to the present invention, the improved VDSR model is trained as follows:
(1) data pre-processing
Selecting a plurality of pictures in the public data set BSD300 as a training set, and selecting a plurality of pictures as a test set;
performing data augmentation (data augmentation) on training data in the training set;
(2) building an improved VDSR model
The improved VDSR model adopts the network structure of the original VDSR model, and replaces the original ReLU function into a novel ReLU function;
(3) training procedure
And (3) inputting the training data in the training set processed in the step (1) into the improved VDSR model built in the step (2) for training to obtain the trained VDSR model.
Further preferably, in step (3), the initial learning rate is set to 0.0001, the optimizer selects Adam, the batch size is set to 16, and 200 epochs are trained, each time after an epoch training is performed.
Further preferably, the data augmentation means: and carrying out random horizontal turnover, random brightness adjustment and random cutting on the training data to expand a training set. Therefore, overfitting of the network is avoided, and the generalization capability of the network is improved.
Preferably, according to the present invention, the output high resolution image uses Peak Signal-to-Noise Ratio (PSNR) as an evaluation index of performance, the Peak PSNR is defined by Mean Square Error (MSE), and assuming that two single-channel images I and K with size m × n, the Mean Square Error is defined as shown in formula (ii):
in the formula (II), I and j respectively represent rows and columns of an image, and I (I, j) and K (I, j) respectively represent pixel values at the ith row and jth column positions of the image I and the image K;
the peak signal-to-noise ratio PSNR is defined as shown in formula (III):
in formula (III), MAXIThe maximum value of a pixel point in the image is represented. If the image is 8-bit, the maximum value is 255.
A computer device comprising a memory storing a computer program and a processor implementing the steps of an image super-resolution method based on an improved VDSR model applying a new ReLU function when executing the computer program.
A computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of an image super-resolution method based on an improved VDSR model applying a novel ReLU function.
The invention has the beneficial effects that:
1. the invention provides a novel ReLU function, which can adaptively learn static working points in a training process and improve the expression capability of a convolutional neural network model.
2. Compared with a VDSR model applying an original ReLU function, the VDSR model applying the novel ReLU function of the invention obtains a higher PSNR value, effectively improves the performance in a super-resolution task, and verifies the effectiveness of the invention.
3. The invention performs data augmentation on the data in the network training process, and adopts a learning rate attenuation strategy, thereby effectively avoiding the overfitting phenomenon in the network training process and improving the generalization capability of the network.
Drawings
Fig. 1 is a schematic diagram of a conventional Sigmoid function image;
FIG. 2 is a diagram of an image of a conventional Tanh function;
FIG. 3 is a diagram of an image of an original ReLU function;
FIG. 4 is a diagram illustrating an image of a ReLU function for adaptively learning a static operating point according to the present invention;
fig. 5 is a schematic diagram of the network architecture of the improved VDSR model of the present invention;
FIG. 6 is a graph showing the result of comparing experimental effects of 2-fold super-resolution tasks;
FIG. 7 is a graph showing the results of comparing the experimental effects of 3-fold super-resolution tasks.
Detailed Description
The invention is further defined in the following, but not limited to, the figures and examples in the description.
Example 1
An image super-resolution method based on an improved VDSR model applying a novel ReLU function is as follows: inputting an image to be processed into a trained improved VDSR model, and outputting to obtain a high-resolution image of the image;
as shown in fig. 4, the new ReLU function in the improved VDSR model, i.e. the ReLU function f (x) of the adaptive learning static operating point, is shown in formula (i), and the abscissa and the ordinate respectively represent the input and the output of the new ReLU function:
in the formula (I), Q refers to a static operating point, the static operating point refers to a zero point in an original ReLU function, Q is obtained through adaptive learning, and x represents the input characteristic of a new ReLU function layer in an improved VDSR model.
Q is obtained through adaptive learning, and means that: setting Q as a learnable parameter and initializing to 0, wherein x represents the input characteristic of a new ReLU function layer in the improved VDSR model, in the training process of the improved VDSR model, the value smaller than Q in the input characteristic is set to 0 by the new ReLU function layer, the value larger than Q is set to x-Q, the Q value is continuously updated in the back propagation process, and after training of a large amount of data, the new ReLU function layers at different positions in the VDSR model have different Q values. Compared with the original ReLU function, the novel ReLU function provided by the invention has better nonlinear fitting capability, and the original ReLU function refers to the ReLU function mentioned in the background art.
The concept of a quiescent operating point stems from the triode amplification principle. The triode has two states, namely a static state and a dynamic state, wherein the static state refers to a direct current working state when no signal is added to the triode, the voltage of each electrode in the static state is called static working current, the dynamic state refers to a working state when an alternating current signal is added to the triode, and the current of each electrode in the dynamic state is called dynamic working current. If the direct current circuit of the triode does not work normally, the alternating current circuit of the triode cannot work normally. The static operating point is that when in a static state, the circuit is in a direct current operating state, and the numerical values of the current and the voltage can be represented by a determined point on an input-output curve of the triode, which is also called as a Q point. The static values of the voltage and the current of the amplifying circuit are determined, and a proper static working point is selected to prevent the circuit from generating nonlinear distortion and ensure the amplifying effect. In the original ReLU function, data with input greater than 0 is output as the original ReLU function, and the input characteristic less than 0 sets the output to 0, and in combination with the concept of the quiescent operating point in the triode, the zero point in the original ReLU function is called the quiescent operating point of the activation function.
As shown in fig. 5, the improved VDSR model includes 20 convolutional layers, 19 new ReLU function layers, and residual concatenation; the improved VDSR model takes the low-resolution image as input after bicubic interpolation, learns the lost high-frequency component of the low-resolution image through 20 layers of convolution layers, and performs pixel-level addition on the learned high-frequency component and the input at the end of the improved VDSR model to obtain the final high-resolution image.
The improved VDSR model is identical in network structure to the original VDSR model, except that the original ReLU function is replaced with the new ReLU function proposed by the present invention.
Example 2
An image super-resolution method based on an improved VDSR model applying a new ReLU function as described in embodiment 1, which is different in that: the training process for the improved VDSR model is as follows:
(1) data pre-processing
200 pictures in a public data set BSD300 are selected as a training set, and 100 pictures are selected as a test set;
performing data augmentation (data augmentation) on training data in the training set;
(2) building an improved VDSR model
The improved VDSR model adopts the network structure of the original VDSR model, and replaces the original ReLU function into a novel ReLU function;
(3) training procedure
And (3) inputting the training data in the training set processed in the step (1) into the improved VDSR model built in the step (2) for training to obtain the trained VDSR model.
In step (3), the initial learning rate is set to 0.0001, the optimizer selects Adam, the batch size is set to 16, 200 epochs are trained, and the test is performed after each epoch training.
Data augmentation refers to: and carrying out random horizontal turnover, random brightness adjustment and random cutting on the training data to expand a training set. Therefore, overfitting of the network is avoided, and the generalization capability of the network is improved.
The output high-resolution image uses Peak Signal-to-Noise Ratio (PSNR) as an evaluation index of performance, the Peak PSNR is defined by Mean Square Error (MSE), and assuming two single-channel images I and K with size m × n, the Mean Square Error is defined as formula (ii):
in the formula (II), I and j respectively represent rows and columns of an image, and I (I, j) and K (I, j) respectively represent pixel values at the ith row and jth column positions of the image I and the image K;
the peak signal-to-noise ratio PSNR is defined as shown in formula (III):
in formula (III), MAXIThe maximum value of a pixel point in the image is represented. If the image is 8-bit, the maximum value is 255.
In this embodiment, 200 pictures in the data set B300 are selected as a training data set, 100 pictures are selected as a testing data set, VDSR models to which the original ReLu function and the novel ReLu function proposed by the present invention are applied are trained respectively,
FIG. 6 is a graph showing the result of comparing experimental effects of 2-fold super-resolution tasks; FIG. 7 is a graph showing the results of comparing the experimental effects of 3-fold super-resolution tasks. In fig. 6 and 7, the abscissa and ordinate respectively represent PSNR (peak signal-to-noise ratio) and epoch (1 epoch represents all samples in 1 training set), the triangle labeled data represents the training result of the improved VDSR model applying the new ReLU function of the present invention, and the circle labeled data represents the training result of the VDSR model applying the original ReLU function;
as can be seen from fig. 6 and 7, in the 2-fold and 3-fold super-resolution tasks, the performance of the VDSR model is improved by applying the novel ReLU function proposed by the present invention.
Example 3
A computer device comprising a memory storing a computer program and a processor implementing the steps of embodiment 1 or 2 of the image super-resolution method based on an improved VDSR model applying a new ReLU function when executing the computer program.
Example 4
A computer-readable storage medium, having stored thereon a computer program which, when executed by a processor, implements the steps of embodiment 1 or 2 of the image super-resolution method based on the improved VDSR model applying the novel ReLU function.
Claims (8)
1. An image super-resolution method based on an improved VDSR model applying a novel ReLU function is characterized in that: inputting an image to be processed into a trained improved VDSR model, and outputting to obtain a high-resolution image of the image;
the new ReLU function in the improved VDSR model, i.e., the ReLU function f (x) of the adaptive learning static operating point, is shown in formula (i):
in the formula (I), Q refers to a static operating point, the static operating point refers to a zero point in an original ReLU function, Q is obtained through adaptive learning, and x represents the input characteristic of a new ReLU function layer in an improved VDSR model.
2. The method of claim 1, wherein the improved VDSR model comprises 20 convolutional layers, 19 novel ReLU function layers, and residual connection; the improved VDSR model takes the low-resolution image as input after bicubic interpolation, learns the lost high-frequency component of the low-resolution image through 20 layers of convolution layers, and performs pixel-level addition on the learned high-frequency component and the input at the end of the improved VDSR model to obtain the final high-resolution image.
3. The method of claim 1, wherein the improved VDSR model is trained as follows:
(1) data pre-processing
Selecting a plurality of pictures in the public data set BSD300 as a training set, and selecting a plurality of pictures as a test set;
performing data augmentation on training data in the training set;
(2) building an improved VDSR model
The improved VDSR model adopts the network structure of the original VDSR model, and replaces the original ReLU function into a novel ReLU function;
(3) training procedure
And (3) inputting the training data in the training set processed in the step (1) into the improved VDSR model built in the step (2) for training to obtain the trained VDSR model.
4. The method of claim 3, wherein in step (3), the initial learning rate is set to 0.0001, the optimizer selects Adam, batchSize is set to 16, 200 epochs are trained, and the testing is performed after each epoch training.
5. The method of claim 1, wherein the data expansion is performed by applying a new ReLU function to the VDSR model: and carrying out random horizontal turnover, random brightness adjustment and random cutting on the training data to expand a training set.
6. The image super-resolution method based on the improved VDSR model applying the new ReLU function as claimed in claim 1, wherein the outputted high resolution image uses peak signal to noise ratio PSNR as an evaluation index of performance, the peak signal to noise ratio PSNR is defined by mean square error MSE, and assuming two single channel images I and K with size m × n, the mean square error is defined as formula (ii):
in the formula (II), I and j respectively represent rows and columns of an image, and I (I, j) and K (I, j) respectively represent pixel values at the ith row and jth column positions of the image I and the image K;
the peak signal-to-noise ratio PSNR is defined as shown in formula (III):
in the formula (III), the compound represented by the formula (III),MAXIthe maximum value of a pixel point in the image is represented.
7. A computer device comprising a memory and a processor, characterized in that the memory stores a computer program which when executed by the processor implements the steps of the method for image super resolution based on an improved VDSR model applying a new ReLU function of any of claims 1-6.
8. A computer-readable storage medium, having stored thereon a computer program, wherein the computer program, when being executed by a processor, is adapted to carry out the steps of the method for image super-resolution based on an improved VDSR model applying a new ReLU function as claimed in any one of claims 1-6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011576889.XA CN112669210B (en) | 2020-12-28 | 2020-12-28 | Image super-resolution method, device and medium based on static working point |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011576889.XA CN112669210B (en) | 2020-12-28 | 2020-12-28 | Image super-resolution method, device and medium based on static working point |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112669210A true CN112669210A (en) | 2021-04-16 |
| CN112669210B CN112669210B (en) | 2022-06-03 |
Family
ID=75410517
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011576889.XA Active CN112669210B (en) | 2020-12-28 | 2020-12-28 | Image super-resolution method, device and medium based on static working point |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112669210B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109102462A (en) * | 2018-08-01 | 2018-12-28 | 中国计量大学 | A kind of video super-resolution method for reconstructing based on deep learning |
| US20190347549A1 (en) * | 2018-05-10 | 2019-11-14 | Microsoft Technology Licensing, Llc | Efficient data encoding for deep neural network training |
| CN110599401A (en) * | 2019-08-19 | 2019-12-20 | 中国科学院电子学研究所 | Remote sensing image super-resolution reconstruction method, processing device and readable storage medium |
| US20200027015A1 (en) * | 2017-04-07 | 2020-01-23 | Intel Corporation | Systems and methods for providing deeply stacked automated program synthesis |
-
2020
- 2020-12-28 CN CN202011576889.XA patent/CN112669210B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200027015A1 (en) * | 2017-04-07 | 2020-01-23 | Intel Corporation | Systems and methods for providing deeply stacked automated program synthesis |
| US20190347549A1 (en) * | 2018-05-10 | 2019-11-14 | Microsoft Technology Licensing, Llc | Efficient data encoding for deep neural network training |
| CN109102462A (en) * | 2018-08-01 | 2018-12-28 | 中国计量大学 | A kind of video super-resolution method for reconstructing based on deep learning |
| CN110599401A (en) * | 2019-08-19 | 2019-12-20 | 中国科学院电子学研究所 | Remote sensing image super-resolution reconstruction method, processing device and readable storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112669210B (en) | 2022-06-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Gai et al. | New image denoising algorithm via improved deep convolutional neural network with perceptive loss | |
| CN106991646B (en) | Image super-resolution method based on dense connection network | |
| US10552944B2 (en) | Image upscaling with controllable noise reduction using a neural network | |
| CN109741364B (en) | Target tracking method and device | |
| CN111932461A (en) | Convolutional neural network-based self-learning image super-resolution reconstruction method and system | |
| CN110648292A (en) | High-noise image denoising method based on deep convolutional network | |
| CN112581397B (en) | Degraded image restoration method, system, medium and equipment based on image priori information | |
| Liu et al. | Learning hadamard-product-propagation for image dehazing and beyond | |
| CN111861886A (en) | Image super-resolution reconstruction method based on multi-scale feedback network | |
| CN111695590A (en) | Deep neural network feature visualization method for constraint optimization class activation mapping | |
| CN111986085A (en) | Image super-resolution method based on depth feedback attention network system | |
| JP6942203B2 (en) | Data processing system and data processing method | |
| CN112669210B (en) | Image super-resolution method, device and medium based on static working point | |
| Chartier et al. | A sequential dynamic heteroassociative memory for multistep pattern recognition and one-to-many association | |
| CN115860113B (en) | Training method and related device for self-countermeasure neural network model | |
| CN116824232A (en) | Data filling type deep neural network image classification model countermeasure training method | |
| CN114092763A (en) | Method for constructing impulse neural network model | |
| Luo et al. | Maximum a posteriori on a submanifold: a general image restoration method with gan | |
| WO2022208632A1 (en) | Inference device, inference method, learning device, learning method, and program | |
| CN117078516B (en) | Mine image super-resolution reconstruction method based on residual mixed attention | |
| CN112949719B (en) | Well testing interpretation proxy model generation method based on GAN | |
| CN111161152B (en) | Image super-resolution method based on self-adaptive convolutional neural network | |
| US20220375032A1 (en) | Image processing apparatus and operating method thereof | |
| CN117522727A (en) | Low-light image processing method based on denoising diffusion probability model | |
| Xiaohua et al. | Total generalized variational-liked network for 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |