CN116594061A - Seismic data denoising method based on multi-scale U-shaped attention network - Google Patents
Seismic data denoising method based on multi-scale U-shaped attention network Download PDFInfo
- Publication number
- CN116594061A CN116594061A CN202310875982.8A CN202310875982A CN116594061A CN 116594061 A CN116594061 A CN 116594061A CN 202310875982 A CN202310875982 A CN 202310875982A CN 116594061 A CN116594061 A CN 116594061A
- Authority
- CN
- China
- Prior art keywords
- module
- attention
- seismic data
- convolution
- shaped
- 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
Landscapes
- Geophysics And Detection Of Objects (AREA)
- Image Processing (AREA)
Abstract
The invention belongs to the field of seismic exploration, in particular to a seismic data denoising method based on a multi-scale U-shaped attention network, which comprises the following steps: the method comprises the steps of inputting collected seismic data into a U-shaped network for data processing, wherein the U-shaped network comprises an upper half part and a lower half part, the upper half part is used for feature extraction, different features are extracted with different resolutions, fine features are extracted in a high resolution stage, and rough features are extracted in a low resolution stage; the lower half part is used for feature fusion, two cascade expansion convolution modules are used on the lowest scale between the upper half part and the lower half part, and the novel network adopted by the invention can extract features on multiple scales. It can obtain more features and local and global information from the seismic data at the same time.
Description
Technical Field
The invention belongs to the field of seismic exploration, and particularly relates to a seismic data denoising method based on a multi-scale U-shaped attention network.
Background
In seismic exploration, the presence of random noise will greatly affect the quality of the seismic data. As random noise becomes more complex, the quality of the acquired field seismic data will continue to degrade. In order to obtain a seismic signal with a high signal-to-noise ratio, it is necessary to denoise the seismic signal to maintain as much effective information as possible. A number of conventional noise suppression methods have been applied to seismic data processing including, for example, wavelet transformation, bandpass filtering, f-x deconvolution, median filtering, curve transformation, empirical mode decomposition, singular value decomposition, and the like. Although some conventional methods can suppress noise in the seismic data to a certain extent, when the seismic wave field is complex, the conventional methods have problems of loss of effective signals, noise residues and the like while removing noise, and the effective signals are often damaged. It is therefore necessary to develop a new effective denoising method.
In recent years, in order to overcome the limitations of the conventional denoising method, a Deep Learning (DL) -based denoising model has been proposed. The DL method solves the parameter problems existing in the conventional method and the complex feature engineering in the classical Machine Learning (ML) technique. And has achieved significant results in the fields of image processing, speech recognition, etc. The method realizes the representation of abstract features step by constructing a plurality of processing layers, so that the inherent rule of data is explored, the method has strong nonlinear mapping expression capability, and the method is widely applied to the seismic exploration process. Such as inversion of the earthquake, reconstruction of the seismic data, suppression of the seismic noise, etc. Convolutional Neural Networks (CNNs), such as DNCNN, are a common deep learning algorithm, a feed-forward neural network with strong feature extraction capability. However, existing CNN-based denoising methods still have shortcomings. The full convolution network is not tightly connected between the shallow layer and the deep layer, which is unfavorable for fully extracting the characteristics. In addition, as the network deepens, the effect of the shallow layer on the deep layer becomes weaker. This results in increased difficulty in network training.
Disclosure of Invention
The invention aims to solve the technical problems of insufficient feature extraction, effective signal loss, noise residue and the like by providing a seismic data denoising method based on a multi-scale U-shaped attention network.
The present invention has been achieved in such a way that,
a method for denoising seismic data based on a multi-scale U-shaped attention network, the method comprising: the method comprises the steps of inputting collected seismic data into a U-shaped network for data processing, wherein the U-shaped network comprises an upper half part and a lower half part, the upper half part is used for feature extraction, different features are extracted with different resolutions, fine features are extracted in a high resolution stage, and rough features are extracted in a low resolution stage; the bottom half is used for feature fusion, and two cascaded dilation convolution modules are used at the lowest scale between the top half and the bottom half.
Further, the upper half includes a first convolution layer and a ReLU activation function to extract local features of the input data and increase the number of channels of the feature map, and the extracted feature map is used as an input of a first residual module, and downsampled after passing through the first residual module.
Further, the downsampling passes through the second residual module and the third residual module to the dilation convolution module.
Further, the lower half part comprises a second convolution layer and a ReLU activation function, the input data is up-sampled, the size of the feature map is increased by two times, the number of channels is reduced by one half, the feature map is up-sampled through a fourth residual error module, the up-sampled data is subjected to feature fusion through the third convolution layer and the ReLU activation function, and finally the feature map is converted into seismic data through a fifth residual error module and then through two fourth convolution and Tanh activation functions.
Further, all residual modules comprise two cascaded attention modules and one convolution layer, wherein the output of the attention module of the first residual module is output to downsampling after one convolution layer, i.e. the size of the feature map is reduced to half of the original, the number of channels is increased to twice of the original, and the second residual module and the third residual module respectively have 128 channels and 256 channels.
Further, the first residual module is in jump connection with the third convolution layer and the ReLU activation function, and the second residual module is in jump connection with the second convolution layer and the ReLU activation function.
Further, the attention module includes two convolution modules that are used to extract features and one channel attention module that is used to make the network extract more useful information, reduce redundant information, and add input features and output features by summing element by element.
Further, the output of the attention module is:in->And->Representing the input and output of the attention module, respectively; />And->A convolution kernel for the first and second convolution layers, respectively; />Is a ReLU activation function; />Is a channel attention module.
Further, the channel attention module comprises an average pooling module and a maximum pooling module, after the data are input into the average pooling module and the maximum pooling module at the same time, the outputs of the maximum pooling module and the average pooling module are input into a weight sharing convolution block, the two outputs are obtained, an attention characteristic diagram is generated through element-by-element summation and addition, the attention weight is between 0 and 1, and the attention weighting and the input characteristic are multiplied element by element.
Compared with the prior art, the invention has the beneficial effects that:
1. traditional deep learning methods only extract features of a single scale. The new network employed by the present invention can extract features on multiple scales. It can obtain more features and local and global information from the seismic data at the same time.
2. The method adopts a jump connection method to connect the shallow layer characteristics and the deep layer characteristics so as to reduce the information loss in the U-shaped network structure. More details can be extracted than in previous methods. In a simulation experiment, the method of the invention obtains the best denoising effect in the simulated seismic data.
3. In the U-shaped network structure, a residual error module consisting of a cascade attention module and a residual error connecting block is provided, and the residual error module is used for replacing a convolution module widely used in a network. The new residual module may make the network more concerned about more important channel information than the normal convolution module. Furthermore, the local residual connection in the attention module also allows bypassing less important information, letting the main network architecture focus on more efficient information. The residual connection block in the residual module also solves the gradient explosion problem which occurs along with the deepening of the network, and further improves the capacity of the network.
Drawings
Fig. 1 is a network structure diagram provided in an embodiment of the present invention;
FIG. 2 is a block diagram of a residual error provided by an embodiment of the present invention;
FIG. 3 is a block diagram of an attention module according to an embodiment of the present invention;
FIG. 4 is a block diagram of a channel attention module according to an embodiment of the present invention;
FIG. 5 shows simulated clean seismic signals and noisy seismic signals provided by embodiments of the invention, (a) first clean seismic data, (b) second clean seismic data, (c) noisy seismic data corresponding to the first clean seismic data, and (d) noisy seismic data corresponding to the second clean seismic data;
fig. 6 is a graph of denoising effects of the methods, where (a) and (e) are median filtering results (b) and (f) are gaussian filtering results (c) and (g) are denoising results (d) and (h) of the DNCNN algorithm, and (h) are denoising results of our network.
Detailed Description
The present invention will be further described in detail with reference to the following examples in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific examples described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
Referring to fig. 1, a seismic data denoising method based on a multi-scale U-shaped attention network, the method comprising: the method comprises the steps of inputting collected seismic data into a U-shaped network for data processing, wherein the U-shaped network comprises an upper half part and a lower half part, the upper half part is used for feature extraction, different features are extracted with different resolutions, fine features are extracted in a high resolution stage, and rough features are extracted in a low resolution stage; the bottom half is used for feature fusion, and two cascaded dilation convolution modules are used at the lowest scale between the top half and the bottom half.
The upper half comprises a first convolution layer and a ReLU activation function 1 to extract local features of the input data and increase the number of channels of the feature map, the extracted feature map is used as input to the first residual module 2, and downsampled after passing through the first residual module 2. Downsampling passes through the second residual block 3 and the third residual block 4 to the dilation convolution block 5.
The lower half part comprises a second convolution layer and a ReLU activation function 6, the input data is up-sampled, the size of a feature map is increased by two times, the number of channels is reduced by one half, the feature map is up-sampled through a fourth residual error module 7, the up-sampled data is subjected to feature fusion through the third convolution layer and the ReLU activation function 8, and finally the feature image is converted into seismic data through a fifth residual error module 9 and then through two fourth convolution and Tanh activation functions 10.
All residual modules comprise two cascaded attention modules and one convolution layer, wherein the output of the attention module of the first residual module is output to downsampling after one convolution layer, i.e. the size of the feature map is reduced to half of the original, the number of channels is increased to twice the original, and the second residual module and the third residual module respectively have 128 and 256 channels.
The first residual module is in jump connection with the third convolution layer and the ReLU activation function, and the second residual module is in jump connection with the second convolution layer and the ReLU activation function.
The upper half of the U-shaped network of the present invention is used for feature extraction. Different features are extracted at different resolutions by downsampling. And the fine features are extracted in the high resolution stage and the coarse features are extracted in the low resolution stage. In the lower half of the U-shaped network for feature fusion. Thus, the structure can learn more data information over multiple scales by downsampling and upsampling. And two cascaded expansion convolution modules are used on the lowest scale of the network, so that the receptive field of the network can be increased without increasing the number of parameters, and the global information can be extracted more comprehensively.
The denoising network adopts a U-shaped structure, and can learn the high-level information and the low-level information of the seismic data at the same time. First, the present invention uses a convolution layer and a ReLU activation function to extract local features of the input data and increase the number of channels of the feature map. The extracted feature map is used as input to the residual module. Residual modules as shown in fig. 2, the residual modules consist of two cascaded attention modules and one convolution layer. The output of the attention module is downsampled after one convolution layer, i.e. the size of the feature map is reduced to half the original and the number of channels is increased to twice the original. Then, two residual modules with 128 and 256 channels are passed, respectively. Two cascaded dilation convolution modules are also used to increase the receptive field of the network at the lowest scale. These enable the network to extract more efficient information without increasing the number of parameters. Next, it is up-sampled by convolution, increasing the size of the feature map by a factor of two, and reducing the number of channels by a factor of half. The convolution and ReLU activation functions are then used to reduce the number of channels of the feature map, which is then passed through the residual module. Similarly, convolution, reLU activation functions, and residual modules are also used in the next scale. And then, performing feature fusion by using convolution and a ReLU activation function, and finally, converting the feature image into seismic data by using the convolution and a Tanh activation function. By comparing with the noise suppression results of several classical denoising methods, the network provided by the invention can effectively keep effective signals, reduce the network depth and better suppress seismic noise. Can be widely applied to the field of seismic data processing.
The downsampling method of the invention can increase the receptive field of the network, so that the network can learn more low-frequency information. The upsampling method may allow the network to learn more high frequency information. And the low-frequency information and the high-frequency information are combined through jump connection, so that noise suppression can be realized while more original information of the seismic data is reserved.
In the example, the invention is verified by constructing data, and for the construction of the dataset, a noisy training set comprising seismic signals and noise signals and a training set comprising only clean seismic signals are established. The invention adopts Rake wavelets to simulate effective signals in seismic records, and the formula is as follows:. A total of 40 synthetic clean seismic signals of size 1000 x 140 are obtained, and then different gaussian white noise is added to the synthetic seismic signals to obtain 40 noisy seismic signals of size 1000 x 140, such as the simulated clean seismic signals and the noisy seismic signals with signal to noise ratios of-4.65 and-7.85 shown in fig. 5. Since a very small patch is detrimental to extracting useful signal features, an excessively large patch increases training costs, sliding windows of patch size 64 x 64 and step size 10 are used to intercept clean and noisy seismic signals, respectively, to generate a signal set containing 23000 clean seismic signal samples and a noisy signal set containing 23000 noisy seismic signal samples, respectively. In the data preprocessing, the invention normalizes each piece of data by using the maximum and minimum values, and the data set is processed according to 7:2:1 is divided into a training set, a verification set and a test set, wherein the training set is used for training a model, the verification set is used for detecting whether the model loss is continuously reduced, and the test set is used for testing the model effect.
All residual modules in the invention have the same structure: as shown in fig. 2, for extracting features in a U-shaped network without having to rely on a plurality of simple convolution modules. The residual module may increase the depth of the network in order to learn more efficient information while avoiding gradient vanishing or gradient explosion. It comprises two cascaded attention modules and one convolution module. The shallow features and the deep features are blended together by means of elemental summation. The output characteristics of the residual module contain more information about the different characteristics
Wherein the attention module: as shown in fig. 3, it consists of two convolution modules and a channel attention module. First, it uses two convolution modules to extract features. Second, the channel attention module is used to make the network extract more useful information and reduce redundant information. Finally, by adding the input features and the output features by element-by-element summation, the network can learn more effective information while retaining the original features. The output of which can be expressed as follows:in->And->Representing the input and output of the attention module, respectively; />And->A convolution kernel for the first and second convolution layers, respectively; />Is a ReLU activation function; />Is a channel attention module.
For the channel attention module: as shown in fig. 4, in order to make the network more concerned about the characteristics containing more relevant information, a channel attention mechanism is used to adaptively adjust the channel characteristics by considering the interrelationship between channels. Firstly, data are simultaneously input into an average pooling module and a maximum pooling module, and input spatial information is integrated with the help of the average pooling module and the maximum pooling module. Second, the outputs of the average pooling module and the maximum pooling module are input to a weight sharing convolution block, and both outputs are obtained. They are added by element-wise summation to generate an attention profile. The attention weight is between 0 and 1. Finally, the attention weighted sum input features are multiplied element by element, which enables features with high information content to be more fully learned.
Qualitative and quantitative experiments are carried out to evaluate and demonstrate the performance of the proposed network model in terms of denoising, three methods of median filtering, gaussian filtering and DNCNN are compared with the method proposed by the invention, and the peak signal-to-noise ratio is used for evaluating the denoising effect, and the formula is as follows:the MSE loss function is used, as follows: />Where m and n represent the size of the data, I represents the network denoising seismic data, and K represents the clean seismic data.
The present invention compares three denoising algorithms, including conventional denoising algorithms such as median filtering, gaussian filtering, and deep learning-based denoising algorithms such as DNCNN algorithm. Fig. 6 shows a denoising effect diagram of each method, wherein (a) and (e) are median filtering results, (b) and (f) are gaussian filtering results, (c) and (g) are denoising results of the DNCNN algorithm, and (d) and (h) are denoising results through the network of the present invention. The specific denoising results are shown in table 1 below.
Table 1 denoising effect evaluation index table
It can be seen from the table that when the noise is smaller or larger, the deep learning-based method is significantly better than the traditional denoising method, and the network of the invention is also significantly better than the classical deep learning-based denoising algorithm, such as DNCNN algorithm. Therefore, the PSNR objective evaluation index is comprehensively compared, and the network provided by the invention is superior to the compared method in denoising performance.
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, and alternatives falling within the spirit and principles of the invention.
Claims (9)
1. A method for denoising seismic data based on a multi-scale U-shaped attention network, the method comprising: the method comprises the steps of inputting collected seismic data into a U-shaped network for data processing, wherein the U-shaped network comprises an upper half part and a lower half part, the upper half part is used for feature extraction, different features are extracted with different resolutions, fine features are extracted in a high resolution stage, and rough features are extracted in a low resolution stage; the bottom half is used for feature fusion, and two cascaded dilation convolution modules are used at the lowest scale between the top half and the bottom half.
2. The method of denoising seismic data based on a multi-scale U-shaped attention network according to claim 1, wherein the upper half comprises a first convolution layer and a ReLU activation function to extract local features of the input data and increase the number of channels of the feature map, the extracted feature map being used as an input to a first residual module through which downsampling is performed.
3. A method of denoising seismic data based on a multiscale U-shaped attention network according to claim 2, wherein the downsampling passes through a second residual block and a third residual block to an expansion convolution block.
4. A method for denoising seismic data based on a multi-scale U-shaped attention network according to claim 3, wherein the lower half comprises a second convolution layer and a ReLU activation function, the input data is upsampled, the size of the feature map is increased by two times, the number of channels is reduced by half, the upsampling of the feature map is completed through a fourth residual module, the upsampled data is subjected to feature fusion through the third convolution layer and the ReLU activation function, and finally the feature map is converted into the seismic data through a fifth residual module and then using two fourth convolution and Tanh activation functions.
5. The method for denoising seismic data based on a multi-scale U-shaped attention network according to claim 4, wherein all residual modules comprise two cascaded attention modules and one convolution layer, wherein the output of the attention module of the first residual module is output to downsampling after one convolution layer, i.e. the size of the feature map is reduced to half, the number of channels is increased to twice the original, and the second residual module and the third residual module have 128 and 256 channels, respectively.
6. The method for denoising seismic data based on a multi-scale U-shaped attention network of claim 4, wherein the first residual module is skip-connected with the third convolution layer and the ReLU activation function, and the second residual module is skip-connected with the second convolution layer and the ReLU activation function.
7. A method of denoising seismic data based on a multi-scale U-shaped attention network according to claim 5 wherein the attention module comprises two convolution modules used to extract features and a channel attention module used to cause the network to extract more useful information, reduce redundant information and add input features and output features by element-wise summation.
8. The multi-scale U-shaped attention-based system of claim 7The seismic data denoising method of the force network is characterized in that the output of the attention module is as follows:in->And->Representing the input and output of the attention module, respectively; />And->A convolution kernel for the first and second convolution layers, respectively; />Is a ReLU activation function; />Is a channel attention module.
9. The method for denoising seismic data based on a multi-scale U-shaped attention network according to claim 7, wherein the channel attention module comprises an average pooling module and a maximum pooling module, the outputs of the maximum pooling module and the average pooling module are input to a weight sharing convolution block after the data are input to the average pooling module and the maximum pooling module at the same time, the two outputs are obtained, and the attention feature map is generated by summing element by element, the attention weight is between 0 and 1, and the attention weighting and the input feature are multiplied element by element.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310875982.8A CN116594061B (en) | 2023-07-18 | 2023-07-18 | Seismic data denoising method based on multi-scale U-shaped attention network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310875982.8A CN116594061B (en) | 2023-07-18 | 2023-07-18 | Seismic data denoising method based on multi-scale U-shaped attention network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN116594061A true CN116594061A (en) | 2023-08-15 |
CN116594061B CN116594061B (en) | 2023-09-22 |
Family
ID=87599528
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310875982.8A Active CN116594061B (en) | 2023-07-18 | 2023-07-18 | Seismic data denoising method based on multi-scale U-shaped attention network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116594061B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117972634A (en) * | 2024-02-08 | 2024-05-03 | 四川鹭羽科技有限公司 | Seismic resolution enhancement method combining attention mechanism and physical mechanism |
Citations (22)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190064378A1 (en) * | 2017-08-25 | 2019-02-28 | Wei D. LIU | Automated Seismic Interpretation Using Fully Convolutional Neural Networks |
US20190064389A1 (en) * | 2017-08-25 | 2019-02-28 | Huseyin Denli | Geophysical Inversion with Convolutional Neural Networks |
CN109903228A (en) * | 2019-02-28 | 2019-06-18 | 合肥工业大学 | A kind of image super-resolution rebuilding method based on convolutional neural networks |
CN111311518A (en) * | 2020-03-04 | 2020-06-19 | 清华大学深圳国际研究生院 | Image denoising method and device based on multi-scale mixed attention residual error network |
CN111833246A (en) * | 2020-06-02 | 2020-10-27 | 天津大学 | Single-frame image super-resolution method based on attention cascade network |
CN112213785A (en) * | 2020-10-19 | 2021-01-12 | 吉林大学 | Seismic data desert noise suppression method based on feature-enhanced denoising network |
US20210287342A1 (en) * | 2020-03-10 | 2021-09-16 | Samsung Electronics Co., Ltd. | Systems and methods for image denoising using deep convolutional networks |
WO2021184620A1 (en) * | 2020-03-19 | 2021-09-23 | 南京昊眼晶睛智能科技有限公司 | Camera-based non-contact heart rate and body temperature measurement method |
CN113592718A (en) * | 2021-08-12 | 2021-11-02 | 中国矿业大学 | Mine image super-resolution reconstruction method and system based on multi-scale residual error network |
CN113935467A (en) * | 2021-10-19 | 2022-01-14 | 吉林大学 | DAS (data acquisition System) borehole exploration data noise suppression method based on iterative multi-scale attention network |
US20220187486A1 (en) * | 2020-12-11 | 2022-06-16 | Hitachi, Ltd. | Computer system and data processing method |
CN114663439A (en) * | 2022-03-21 | 2022-06-24 | 中国人民解放军战略支援部队信息工程大学 | Remote sensing image land and sea segmentation method |
US20220221609A1 (en) * | 2021-01-14 | 2022-07-14 | Cgg Services Sas | SEISMIC DATA PROCESSING USING DUnet |
CN115170392A (en) * | 2022-06-23 | 2022-10-11 | 东南大学 | Single-image super-resolution algorithm based on attention mechanism |
CN115375711A (en) * | 2022-09-19 | 2022-11-22 | 安徽大学 | Image segmentation method of global context attention network based on multi-scale fusion |
WO2022242029A1 (en) * | 2021-05-18 | 2022-11-24 | 广东奥普特科技股份有限公司 | Generation method, system and apparatus capable of visual resolution enhancement, and storage medium |
CN115424103A (en) * | 2022-08-18 | 2022-12-02 | 重庆理工大学 | Improved U-Net brain tumor segmentation method based on attention mechanism and multi-scale feature fusion |
CN115601750A (en) * | 2022-09-14 | 2023-01-13 | 河北地质大学(Cn) | Seismic facies recognition semantic segmentation method and system for improving edge accuracy |
CN115877461A (en) * | 2022-12-23 | 2023-03-31 | 吉林大学 | Desert earthquake noise suppression method based on multi-scale attention interaction network |
CN115984631A (en) * | 2023-02-15 | 2023-04-18 | 西安交通大学 | Markless cell transmission light microscopic image virtual staining method based on deep learning automatic optimization |
CN116184502A (en) * | 2022-12-02 | 2023-05-30 | 吉林大学 | Underground DAS noise suppression method based on subspace projection attention network |
CN116246169A (en) * | 2023-02-13 | 2023-06-09 | 成都理工大学 | SAH-Unet-based high-resolution remote sensing image impervious surface extraction method |
-
2023
- 2023-07-18 CN CN202310875982.8A patent/CN116594061B/en active Active
Patent Citations (22)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190064378A1 (en) * | 2017-08-25 | 2019-02-28 | Wei D. LIU | Automated Seismic Interpretation Using Fully Convolutional Neural Networks |
US20190064389A1 (en) * | 2017-08-25 | 2019-02-28 | Huseyin Denli | Geophysical Inversion with Convolutional Neural Networks |
CN109903228A (en) * | 2019-02-28 | 2019-06-18 | 合肥工业大学 | A kind of image super-resolution rebuilding method based on convolutional neural networks |
CN111311518A (en) * | 2020-03-04 | 2020-06-19 | 清华大学深圳国际研究生院 | Image denoising method and device based on multi-scale mixed attention residual error network |
US20210287342A1 (en) * | 2020-03-10 | 2021-09-16 | Samsung Electronics Co., Ltd. | Systems and methods for image denoising using deep convolutional networks |
WO2021184620A1 (en) * | 2020-03-19 | 2021-09-23 | 南京昊眼晶睛智能科技有限公司 | Camera-based non-contact heart rate and body temperature measurement method |
CN111833246A (en) * | 2020-06-02 | 2020-10-27 | 天津大学 | Single-frame image super-resolution method based on attention cascade network |
CN112213785A (en) * | 2020-10-19 | 2021-01-12 | 吉林大学 | Seismic data desert noise suppression method based on feature-enhanced denoising network |
US20220187486A1 (en) * | 2020-12-11 | 2022-06-16 | Hitachi, Ltd. | Computer system and data processing method |
US20220221609A1 (en) * | 2021-01-14 | 2022-07-14 | Cgg Services Sas | SEISMIC DATA PROCESSING USING DUnet |
WO2022242029A1 (en) * | 2021-05-18 | 2022-11-24 | 广东奥普特科技股份有限公司 | Generation method, system and apparatus capable of visual resolution enhancement, and storage medium |
CN113592718A (en) * | 2021-08-12 | 2021-11-02 | 中国矿业大学 | Mine image super-resolution reconstruction method and system based on multi-scale residual error network |
CN113935467A (en) * | 2021-10-19 | 2022-01-14 | 吉林大学 | DAS (data acquisition System) borehole exploration data noise suppression method based on iterative multi-scale attention network |
CN114663439A (en) * | 2022-03-21 | 2022-06-24 | 中国人民解放军战略支援部队信息工程大学 | Remote sensing image land and sea segmentation method |
CN115170392A (en) * | 2022-06-23 | 2022-10-11 | 东南大学 | Single-image super-resolution algorithm based on attention mechanism |
CN115424103A (en) * | 2022-08-18 | 2022-12-02 | 重庆理工大学 | Improved U-Net brain tumor segmentation method based on attention mechanism and multi-scale feature fusion |
CN115601750A (en) * | 2022-09-14 | 2023-01-13 | 河北地质大学(Cn) | Seismic facies recognition semantic segmentation method and system for improving edge accuracy |
CN115375711A (en) * | 2022-09-19 | 2022-11-22 | 安徽大学 | Image segmentation method of global context attention network based on multi-scale fusion |
CN116184502A (en) * | 2022-12-02 | 2023-05-30 | 吉林大学 | Underground DAS noise suppression method based on subspace projection attention network |
CN115877461A (en) * | 2022-12-23 | 2023-03-31 | 吉林大学 | Desert earthquake noise suppression method based on multi-scale attention interaction network |
CN116246169A (en) * | 2023-02-13 | 2023-06-09 | 成都理工大学 | SAH-Unet-based high-resolution remote sensing image impervious surface extraction method |
CN115984631A (en) * | 2023-02-15 | 2023-04-18 | 西安交通大学 | Markless cell transmission light microscopic image virtual staining method based on deep learning automatic optimization |
Non-Patent Citations (3)
Title |
---|
崔少华;李素文;汪徐德;: "改进的卷积神经网络对地震数据进行去噪的方法", 激光与光电子学进展, no. 06 * |
张攀龙;李尧;张田涛;岳景杭;董锐;曹帅;张庆松;: "基于U-Net深度神经网络的地震数据去噪研究", 金属矿山, no. 01 * |
罗仁泽;李阳阳;: "一种基于RUnet卷积神经网络的地震资料随机噪声压制方法", 石油物探, no. 01 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117972634A (en) * | 2024-02-08 | 2024-05-03 | 四川鹭羽科技有限公司 | Seismic resolution enhancement method combining attention mechanism and physical mechanism |
Also Published As
Publication number | Publication date |
---|---|
CN116594061B (en) | 2023-09-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112507997B (en) | Face super-resolution system based on multi-scale convolution and receptive field feature fusion | |
CN112233026A (en) | SAR image denoising method based on multi-scale residual attention network | |
CN111028163A (en) | Convolution neural network-based combined image denoising and weak light enhancement method | |
CN111340046A (en) | Visual saliency detection method based on feature pyramid network and channel attention | |
CN111415323B (en) | Image detection method and device and neural network training method and device | |
CN111127354B (en) | Single-image rain removing method based on multi-scale dictionary learning | |
Liu et al. | The classification and denoising of image noise based on deep neural networks | |
CN116594061B (en) | Seismic data denoising method based on multi-scale U-shaped attention network | |
CN111598804B (en) | Deep learning-based image multi-level denoising method | |
CN112435162B (en) | Terahertz image super-resolution reconstruction method based on complex domain neural network | |
CN111915486B (en) | Confrontation sample defense method based on image super-resolution reconstruction | |
CN108062575A (en) | A kind of high similarity graph picture identification and sorting technique | |
CN113865859A (en) | Multi-scale multi-source heterogeneous information fusion gearbox state fault diagnosis method | |
CN112634159B (en) | Hyperspectral image denoising method based on blind noise estimation | |
CN108830829B (en) | Non-reference quality evaluation algorithm combining multiple edge detection operators | |
Li et al. | Densely connected network for impulse noise removal | |
CN115877461A (en) | Desert earthquake noise suppression method based on multi-scale attention interaction network | |
CN113962882B (en) | JPEG image compression artifact eliminating method based on controllable pyramid wavelet network | |
CN115205136A (en) | Image rain removing method based on Fourier prior | |
CN111047537A (en) | System for recovering details in image denoising | |
CN113362241B (en) | Depth map denoising method combining high-low frequency decomposition and two-stage fusion strategy | |
CN113484913A (en) | Seismic data denoising method with multi-granularity feature fusion convolution neural network | |
CN114331853A (en) | Single image restoration iteration framework based on target vector updating module | |
CN113361417B (en) | Human behavior recognition method based on variable time sequence | |
CN113781340A (en) | Depth network model for synthetic aperture radar image speckle suppression |
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 |