CN114511473A - Hyperspectral remote sensing image denoising method based on unsupervised adaptive learning - Google Patents
Hyperspectral remote sensing image denoising method based on unsupervised adaptive learning Download PDFInfo
- Publication number
- CN114511473A CN114511473A CN202210409199.8A CN202210409199A CN114511473A CN 114511473 A CN114511473 A CN 114511473A CN 202210409199 A CN202210409199 A CN 202210409199A CN 114511473 A CN114511473 A CN 114511473A
- Authority
- CN
- China
- Prior art keywords
- denoising
- image
- network
- discriminator
- loss
- 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
- 230000003044 adaptive effect Effects 0.000 title claims abstract description 32
- 238000000034 method Methods 0.000 title claims abstract description 32
- 238000001228 spectrum Methods 0.000 claims abstract description 18
- 238000012549 training Methods 0.000 claims abstract description 8
- 230000006870 function Effects 0.000 claims description 24
- 230000004913 activation Effects 0.000 claims description 18
- 238000004364 calculation method Methods 0.000 claims description 13
- 238000010606 normalization Methods 0.000 claims description 9
- 238000004422 calculation algorithm Methods 0.000 claims description 6
- 238000013507 mapping Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 238000010586 diagram Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 239000011159 matrix material Substances 0.000 claims description 3
- 239000000126 substance Substances 0.000 claims description 2
- 230000015556 catabolic process Effects 0.000 abstract description 3
- 238000013135 deep learning Methods 0.000 abstract description 3
- 238000013136 deep learning model Methods 0.000 abstract description 3
- 238000006731 degradation reaction Methods 0.000 abstract description 3
- 238000004088 simulation Methods 0.000 abstract description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000011176 pooling Methods 0.000 description 2
- 230000005855 radiation Effects 0.000 description 2
- 238000007792 addition Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A40/00—Adaptation technologies in agriculture, forestry, livestock or agroalimentary production
- Y02A40/10—Adaptation technologies in agriculture, forestry, livestock or agroalimentary production in agriculture
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention relates to a hyperspectral remote sensing image denoising method based on unsupervised adaptive learning. Aiming at the problem that the generalization of the model is reduced by the degradation difference between the simulated image and the real image, the unsupervised self-adaptive learning strategy is provided, the pre-training is carried out on the high-quality ground image, the discriminator is designed to carry out modeling on noise, the discriminator is used for carrying out fine adjustment on the denoising parameter when the real image is processed, and the generalization of the model on the real image is improved. According to the invention, a depth denoising network based on space-spectrum residual errors and a discriminator based on global information modeling are designed in a model so as to fully excavate hyperspectral depth prior. The method can solve the problem of simulation training data in the hyperspectral remote sensing image deep learning denoising, reduces the dependence of a deep learning model on the simulation training data, and effectively improves the applicability and precision of hyperspectral denoising.
Description
Technical Field
The invention relates to the technical field of remote sensing image processing, in particular to a hyperspectral remote sensing image denoising method based on unsupervised adaptive learning.
Background
With the increase of the spectral resolution of the sensor, the hyperspectral remote sensing is rapidly developed. The hyperspectral satellite can quickly acquire images of hundreds of wave bands in the same area, and has the advantage of integrating maps. However, in the process of acquiring data by the hyperspectral remote sensing satellite, due to the bad atmospheric conditions, the imperfect correction process and the defects of the sensor, the hyperspectral remote sensing image is inevitably polluted by various noises, so that the degradation problems of stripe noise, random noise, image contrast reduction and the like are generated, and the usability of the image data is severely limited. In the image data containing mixed noise, the image information consists of original radiation information reflecting ground real characteristics and noise information, and the noise seriously reduces the image quality and precision. Therefore, it is vital to remove mixed noise in the hyperspectral remote sensing satellite image and truly restore image radiation information, and the removal of the mixed noise can greatly improve the usability of the hyperspectral remote sensing satellite image data and is beneficial to the subsequent processing of the image, such as resource exploration, change detection, land utilization type investigation and the like.
At present, the existing hyperspectral mixed noise removal method has great limitation in practical application. The filtering-based method is limited by a fixed transform domain, the geometric characteristics of hyperspectral data are difficult to fully mine, and a proper cut-off frequency is difficult to determine in practical application; the regularization model-based method is based on artificial assumptions on noise characteristics and potential noise-free image characteristics, accurate modeling of complex noise in massive hyperspectral data is difficult, and when more constraint terms are added, the solving process is very complex, the operation efficiency is low, and the practical application requirements are difficult to meet; the deep learning-based method has high-efficiency operation efficiency, but the training process of the deep learning-based method depends on a large number of noise-free image pairs, because airborne and spaceborne hyperspectral remote sensing platforms can not obtain real image pairs generally, deep learning models are mostly trained on simulated image pairs, and the generalization of the models on real data is reduced. Therefore, the deep learning model capable of improving the generalization of the model and efficiently completing the task of removing the mixed noise of the hyperspectral remote sensing image in a robust mode is researched, the quality of the existing satellite data is improved, and beneficial guidance is provided for efficient preprocessing of the hyperspectral remote sensing satellite.
Disclosure of Invention
The invention aims to provide a hyperspectral remote sensing image denoising method based on unsupervised adaptive learning.
The hyperspectral remote sensing image denoising method based on unsupervised adaptive learning provided by the invention provides an unsupervised adaptive learning strategy aiming at the problem that the model generalization is reduced by the degradation difference between the simulation data and the real data, pre-trains on a high-quality ground image, designs a discriminator to model noise, and finely adjusts denoising parameters by the discriminator when the real data is processed, thereby improving the model generalization. According to the invention, a depth denoising network based on space-spectrum residual errors and a discriminator based on global information modeling are designed in a model so as to fully excavate hyperspectral depth prior.
The hyperspectral remote sensing image denoising method based on unsupervised adaptive learning provided by the invention has the following three remarkable characteristics. Firstly, a training strategy based on unsupervised adaptive learning enables a model to learn a common image prior of a simulated image and can finely adjust denoising parameters according to a real image, so that the generalization of the model is improved; secondly, a discriminator is innovatively introduced to learn the noise mode, unsupervised learning on a real image is realized, and the global space-spectrum modeling capability of the model is effectively improved through global pooling; and thirdly, aiming at the characteristics of the hyperspectral denoising task, a space-spectrum residual volume block is designed to fully excavate the bottom layer characteristics in the model, so that the spatial detail information in the denoising result is more accurately recovered.
The invention provides a hyperspectral remote sensing image denoising method based on unsupervised adaptive learning, which comprises the following implementation steps of:
step 1, inputting a ground hyperspectral image
Normalizing the image data to make the pixel values distributed in the range of 0-1, and then locating the pixel valuesForming a block of image of a certain size, adding analog noise to the block to generate a degraded image
Wherein, in the step (A),H，W，Bthe number of image rows, columns and wave bands respectively;
step 2, the initial parameters of the de-noising network are adjusted
And discriminator initial parameter
Carrying out initialization;
step 3, training a denoising network and a discriminator based on an algorithm of adaptive moment estimation, and setting the current iteration times ask；
Step 3.1, simulating the degraded image
Inputting the data into a denoising network, and outputting a denoising result by the denoising network through calculation
Wherein, in the step (A),
is as followskThe mapping function of the de-noising network at the time of the sub-iteration,
is as followskDenoising parameters of the network during the secondary iteration;
step 3.2, respectively using the noiseless images
And denoising results
Inputting the image into a discriminator, and calculating the probability that the current input is a noise-free image
Corresponding probability value of
De-noising result
Corresponding probability value of
Wherein, in the step (A),
is as followskThe mapping function of the discriminator at the time of the sub-iteration,
is as followskParameters of the discriminator during the secondary iteration;
step 3.3, calculating the loss of the denoising network, including MSE loss
And fight against loss
；
Step 3.4, calculating the discrimination loss of the discriminator
；
WhereinnFor the set batch size,
represents the L2 norm;
step 3.5, loss according to MSE
And fight against loss
Calculating gradient, updating denoising network parameters by back propagation, and outputtingk + Parameters of denoised network at 1 iteration
According to the judgment of the loss
Calculating gradient, back-propagating and updating discriminator parameters, and outputtingk + Parameters of the discriminator at 1 iteration
；
Step 3.6, judging whether the current iteration times exceed a certain number, if not, performing step 3.1, and if so, performing step 4;
step 4, inputting a real hyperspectral image
Normalizing the image data to make the pixel value distribution in the range of 0 to 1,
，
，
the number of real image lines, columns and wave bands respectively;
step 5, carrying out algorithm based on adaptive moment estimation on the trained denoising network
Fine-tuning is performed, wherein,
setting the current iteration times as
；
Step 6, the real degraded image is processed
Input to post-trim denoising network
In the method, the denoising network outputs a final denoising result through calculation
Wherein, in the step (A),
and 5, outputting the denoised network parameters output in the step 5.
Further, the analog noise in step 1 includes gaussian noise
A strip noise and an impulse noise,
is the standard deviation of gaussian noise.
Further, in the step 2, a He initialization-based mode is adopted to carry out initial parameters on the denoising network
And discriminator initial parameter
Initialization is performed.
Further, in step 2, the layer 1 of the denoising network is composed of a convolution layer Conv and a Parametric ReLU activation function, the layers 2-13 are space-spectrum residual convolution blocks D-block, the layer 14 is composed of a convolution layer Conv and a batch normalization layer BN, a feature map of a space-spectrum residual convolution module is output, the layers 15-16 are composed of a convolution layer Conv, a batch normalization layer BN and a Parametric ReLU activation function, finally, a denoising result is output through convolution, and the space-spectrum residual convolution block and the input and the output of the denoising network are connected by using a skip layer; the 1 st layer of the discriminator consists of convolution layer Conv and Leaky ReLU activation functions, the 2 nd to 8 th layers are LeakyConv and consist of convolution layer Conv, batch normalization layer BN and Leaky ReLU activation functions, the step length of the LeakyConv of the even layer is set to be 2, and the 9 th layer outputs a probability matrix through the convolution layer Conv and Sigmoid activation functions;
the space-spectrum residual error convolution block is formed by two layers of multi-channel two-dimensional convolution and is followed by a batch normalization layer BN, the first layer BN is followed by a Parametric ReLU activation function, and meanwhile, a feature graph output by the D-block and an input are added by using layer skipping connection.
Further, loss of MSE
And fight against loss
The specific calculation formula of (A) is as follows;
wherein the content of the first and second substances,nis the set batch size.
Further, the specific implementation manner of step 5 is as follows;
step 5.1, real degraded image
Inputting the data into a denoising network, and outputting a denoising result by the denoising network through calculation
Wherein, in the process,
is as follows
Denoising parameters of the network during the secondary iteration;
step 5.2, denoising the noise-free image
Inputting the noise-free image into a discriminator, and calculating the probability that the current denoising result is a noiseless image by the discriminator
Wherein, in the step (A),
the discriminator parameter outputted in the step 3;
step 5.3, calculating the loss of the de-noising network, including the consistency constraint loss
And unsupervised adaptive loss
；
Step 5.4, loss constraint according to consistency
And unsupervised adaptive loss
Calculating gradient, back-propagating to update de-noising network parameters, and outputting
Parameters of denoised network at sub-iteration
；
And 5.5, judging whether the current iteration times exceed a certain number, if not, performing the step 5.1, and if so, performing the step 6.
Further, consistency constraint loss
And unsupervised adaptive loss
The specific calculation formula of (A) is as follows;
the method of the invention has the following remarkable effects: (1) after the analog image general image prior is learned, the real image prior can be further learned, and the model generalization is improved; (2) a discriminator is designed to learn the noise mode in the image, unsupervised learning on the real degraded image is realized, and the global information modeling capability of the model is further improved by using global pooling; (3) and designing a space-spectrum residual convolution block to ensure that the network utilizes the bottom layer characteristics and accurately recovers the spatial detail information.
Drawings
FIG. 1 is an overall flow diagram of an embodiment of the present invention.
Fig. 2 is a structural diagram of a denoising network, a discriminator and a spatio-spectral residual convolution block in step 2 according to the embodiment of the present invention.
Fig. 3 is a final hyperspectral image denoising result output in step 6 in the embodiment of the present invention.
Detailed Description
The technical scheme of the invention is further specifically described by the following embodiments and the accompanying drawings.
As shown in FIG. 1, the hyperspectral remote sensing image denoising method based on unsupervised adaptive learning provided by the invention comprises the following steps:
step 1, inputting 100 ground hyperspectral images from an ICVL data set
Normalizing the image data to make the pixel value distribution in the range of 0-1, processing the pixel value distribution into 40 x 10 image blocks, and adding analog noise to the image blocks to generate a degraded image
The analog noise includes Gaussian noise
A strip noise and an impulse noise, wherein,H，W，Bthe number of image rows, columns and bands,
is the standard deviation of gaussian noise; considering the difference of the noise intensity of each wave band of the real hyperspectral image, randomly selecting the simulated noise intensity of each wave band, wherein the standard deviation of Gaussian noise
Standard deviation of band noise
Percentage of impulse noise
。
Step 2, carrying out initial parameter pair on the denoising network based on He initialization mode
And discriminator initial parameter
Initializing, wherein a denoising network, a discriminator and a space-spectrum residual volume block (D-block) in the denoising network are shown in FIG. 2; the empty-spectrum residual volume block is composed of two layers of multichannel two-dimensional convolution (Conv), a Batch Normalization Layer (BN) is arranged behind the empty-spectrum residual volume block, a Parametric ReLU (PReLU) activation function is arranged behind the first Layer BN, and meanwhile, the feature graph of the D-block output and the input are added by using Layer skip connection; the layer 1 of the denoising network is composed of Conv and a PReLU activation function, the layers 2-13 are D-blocks, the layer 14 is composed of Conv and BN, a characteristic diagram of an empty-spectrum residual convolution module is output, the layers 15-16 are composed of Conv, BN and a PReLU activation function, finally, a denoising result is output through convolution, and the empty-spectrum residual convolution block and the input and the output of the denoising network are connected by using skip layers; the 1 st layer of the discriminator consists of Conv and Leaky ReLU activation functions, the 2 nd to 8 th layers are LeakyConv which consists of Conv, BN and Leaky ReLU activation functions, the step size of the LeakyConv at the even layer is set to be 2, and the 9 th layer outputs a probability matrix through the Conv and Sigmoid activation functions.
Step 3, training a denoising network and a discriminator based on an algorithm of adaptive moment estimation, setting the batch size to be 128, and setting the current iteration times to bek；
Step 3.1, simulating the degraded image
Inputting the data into a denoising network, and outputting a denoising result by the denoising network through calculation
Wherein, in the step (A),
is as followskThe mapping function of the de-noising network at the time of the sub-iteration,
is as followskDenoising parameters of the network during the secondary iteration;
step 3.2, respectively using the noiseless images
And denoising results
Inputting the image into a discriminator, and calculating the probability that the current input is a noise-free image
Corresponding probability value of
De-noising result
Corresponding probability value of
Wherein, in the step (A),
is as followskThe mapping function of the discriminator at the time of the sub-iteration,
is as followskParameters of the discriminator during the secondary iteration;
step 3.3, calculating the loss of the denoising network, including MSE loss
And fight against loss
Wherein, in the step (A),nis the set batch size;
step 3.4, calculating the discrimination loss of the discriminator
；
Step 3.5, loss according to MSE
And fight against loss
Calculating gradient, back-propagating to update de-noising network parameters, and outputtingk + Parameters of denoised network at 1 iteration
According to the judgment of the loss
Calculating gradient, back-propagating and updating discriminator parameters, and outputtingk + Parameters of the arbiter at 1 iteration
；
And 3.6, judging whether the current iteration times exceed 500 times, if not, performing the step 3.1, and if so, performing the step 4.
Step 4, inputting the hyperspectral image of the WHU-Hi-Baoxie unmanned aerial vehicle
Normalizing the image data to make the pixel value distribution in the range of 0 to 1, wherein,
，
，
the number of real image rows, columns and bands.
Step 5, carrying out algorithm based on adaptive moment estimation on the trained denoising network
Fine-tuning is performed, wherein,
setting the current iteration times as
；
Step 5.1, real degraded image
Inputting the data into a denoising network, and outputting a denoising result by the denoising network through calculation
Wherein, in the step (A),
is as follows
Denoising parameters of the network during the secondary iteration;
step 5.2, denoising the noise-free image
Inputting the noise-free image into a discriminator, and calculating the probability that the current denoising result is a noiseless image by the discriminator
Wherein, in the step (A),
the discriminator parameter outputted in the step 3;
step 5.3, calculating the loss of the denoising network, including consistency constraint loss
And unsupervised adaptive loss
；
Step 5.4, loss constraint according to consistency
And unsupervised adaptive loss
Calculating gradient, updating denoising network parameters by back propagation, and outputting
Parameters of de-noising network at sub-iteration
；
And 5.5, judging whether the current iteration times exceed 10 times, if not, performing the step 5.1, and if so, performing the step 6.
Step 6, the real degraded image is processed
Input to post-trim denoising network
In the method, the denoising network outputs a final denoising result of the hyperspectral image of the WHU-Hi-Baoxie unmanned aerial vehicle through calculation
As shown in fig. 3, in which,
and 5, outputting the denoised network parameters output in the step 5.
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.
Claims (7)
1. A hyperspectral remote sensing image denoising method based on unsupervised adaptive learning is characterized by comprising the following steps:
step 1, inputting a ground hyperspectral image
Normalizing the image data to make the pixel values distributed in the range of 0-1, processing the pixel values into image blocks with a certain size, and adding analog noise to the image blocks to generate a degraded image
Wherein, in the step (A),H，W，Bthe number of image rows, columns and wave bands respectively;
step 2, carrying out initial parameters on the denoising network
And discriminator initial parameter
Carrying out initialization;
step 3, training a denoising network and a discriminator based on an algorithm of adaptive moment estimation, and setting the current iteration times ask；
Step 3.1, simulating the degraded image
Inputting the data into a denoising network, and outputting a denoising result by the denoising network through calculation
Wherein, in the process,
is as followskThe mapping function of the de-noising network at the time of the sub-iteration,
is a firstkDenoising parameters of the network during the secondary iteration;
step 3.2, respectively using the noiseless images
And denoising results
Inputting the image into a discriminator, and calculating the probability that the current input is a noise-free image
Corresponding probability value of
De-noising result
Corresponding probability value of
Wherein, in the step (A),
is as followskThe mapping function of the discriminator at the time of the sub-iteration,
is as followskParameters of the discriminator during the secondary iteration;
step 3.3, calculating the loss of the denoising network, including MSE loss
And to combat the loss
；
Step 3.4, calculating the discrimination loss of the discriminator
；
WhereinnFor the set batch size to be set,
represents the L2 norm;
step 3.5, loss according to MSE
And fight against loss
Calculating gradient, back-propagating to update de-noising network parameters, and outputtingk + De-noising at 1 iterationParameters of a network
According to the judgment of the loss
Calculating gradient, back-propagating and updating discriminator parameters, and outputtingk + Parameters of the arbiter at 1 iteration
；
Step 3.6, judging whether the current iteration times exceed a certain number, if not, performing step 3.1, and if so, performing step 4;
step 4, inputting a real hyperspectral image
Normalizing the image data to make the pixel value distribution in the range of 0 to 1, wherein,
，
，
the number of real image lines, columns and wave bands respectively;
step 5, carrying out algorithm based on adaptive moment estimation on the trained denoising network
Fine-tuning is performed, wherein,
setting the current iteration times as
；
Step 6, the real degraded image is processed
Input to post-trim denoising network
In the method, the denoising network outputs a final denoising result through calculation
Wherein, in the step (A),
and 5, outputting the denoised network parameters output in the step 5.
2. The hyperspectral remote sensing image denoising method based on unsupervised adaptive learning of claim 1, wherein: the analog noise in step 1 includes gaussian noise
A strip noise and an impulse noise,
is the standard deviation of gaussian noise.
3. The hyperspectral remote sensing image denoising method based on unsupervised adaptive learning of claim 1, wherein: step 2, adopting a He-based initialization mode to carry out initial parameters on the denoising network
And discriminator initial parameter
Initialization is performed.
4. The hyperspectral remote sensing image denoising method based on unsupervised adaptive learning of claim 1, wherein: in the step 2, the layer 1 of the denoising network consists of a convolution layer Conv and a Parametric ReLU activation function, the layers 2-13 are space-spectrum residual convolution blocks D-block, the layer 14 consists of the convolution layer Conv and a batch normalization layer BN, a characteristic diagram of a space-spectrum residual convolution module is output, the layers 15-16 consist of the convolution layer Conv, the batch normalization layer BN and the Parametric ReLU activation function, finally, a denoising result is output through convolution, and the input and the output of the space-spectrum residual convolution blocks and the denoising network are connected by using skip layers; the 1 st layer of the discriminator consists of convolution layer Conv and Leaky ReLU activation functions, the 2 nd to 8 th layers are LeakyConv and consist of convolution layer Conv, batch normalization layer BN and Leaky ReLU activation functions, the step length of the LeakyConv of the even layer is set to be 2, and the 9 th layer outputs a probability matrix through the convolution layer Conv and Sigmoid activation functions;
the space-spectrum residual error convolution block is formed by two layers of multi-channel two-dimensional convolution and is followed by a batch normalization layer BN, the first layer BN is followed by a Parametric ReLU activation function, and meanwhile, a feature graph output by the D-block and an input are added by using layer skipping connection.
5. The hyperspectral remote sensing image denoising method based on unsupervised adaptive learning of claim 1, wherein: loss of MSE
And fight against loss
The specific calculation formula of (A) is as follows;
wherein the content of the first and second substances,nis the set batch size.
6. The hyperspectral remote sensing image denoising method based on unsupervised adaptive learning of claim 1, wherein: the specific implementation manner of the step 5 is as follows;
step 5.1, real degraded image
Inputting the data into a denoising network, and outputting a denoising result by the denoising network through calculation
Wherein, in the step (A),
is as follows
Denoising parameters of the network during the secondary iteration;
step 5.2, denoising the noise-free image
Inputting the noise-free image into a discriminator, and calculating the probability that the current denoising result is a noiseless image by the discriminator
Wherein, in the step (A),
the discriminator parameter outputted in the step 3;
step 5.3, calculating the loss of the de-noising network, including the consistency constraint loss
And unsupervised adaptive loss
；
Step 5.4, loss constraint according to consistency
And unsupervised adaptive loss
Calculating gradient, back-propagating to update de-noising network parameters, and outputting
Parameters of de-noising network at sub-iteration
；
And 5.5, judging whether the current iteration times exceed a certain number, if not, performing the step 5.1, and if so, performing the step 6.
7. The hyperspectral remote sensing image denoising method based on unsupervised adaptive learning of claim 6, wherein: loss of consistency constraint
And unsupervised adaptive loss
The specific calculation formula of (A) is as follows;
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210409199.8A CN114511473B (en) | 2022-04-19 | 2022-04-19 | Hyperspectral remote sensing image denoising method based on unsupervised adaptive learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210409199.8A CN114511473B (en) | 2022-04-19 | 2022-04-19 | Hyperspectral remote sensing image denoising method based on unsupervised adaptive learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114511473A true CN114511473A (en) | 2022-05-17 |
| CN114511473B CN114511473B (en) | 2022-07-05 |
Family
ID=81555535
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210409199.8A Active CN114511473B (en) | 2022-04-19 | 2022-04-19 | Hyperspectral remote sensing image denoising method based on unsupervised adaptive learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114511473B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115546653A (en) * | 2022-12-02 | 2022-12-30 | 耕宇牧星(北京)空间科技有限公司 | Remote sensing image denoising method based on depth enhancement network |
| CN116071265A (en) * | 2023-02-21 | 2023-05-05 | 哈尔滨工业大学 | Image denoising method based on space self-adaptive self-supervision learning |
| CN116861167A (en) * | 2023-06-12 | 2023-10-10 | 河北工程大学 | FBG spectrum cyclic denoising method based on deep learning |
| CN117876692A (en) * | 2024-03-11 | 2024-04-12 | 中国石油大学(华东) | Feature weighted connection guided single-image remote sensing image denoising method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110320162A (en) * | 2019-05-20 | 2019-10-11 | 广东省智能制造研究所 | A kind of semi-supervised high-spectral data quantitative analysis method based on generation confrontation network |
| CN110991636A (en) * | 2019-11-14 | 2020-04-10 | 东软医疗系统股份有限公司 | Training method and device of generative confrontation network, image enhancement method and equipment |
| US20210304364A1 (en) * | 2020-03-26 | 2021-09-30 | Bloomberg Finance L.P. | Method and system for removing noise in documents for image processing |
| CN113808042A (en) * | 2021-09-16 | 2021-12-17 | 沈阳工业大学 | SAR image denoising method based on wavelet transformation and generation countermeasure network |
| WO2022041521A1 (en) * | 2020-08-31 | 2022-03-03 | 浙江大学 | Low-dose sinogram denoising and pet image reconstruction method based on teacher-student generators |
-
2022
- 2022-04-19 CN CN202210409199.8A patent/CN114511473B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110320162A (en) * | 2019-05-20 | 2019-10-11 | 广东省智能制造研究所 | A kind of semi-supervised high-spectral data quantitative analysis method based on generation confrontation network |
| CN110991636A (en) * | 2019-11-14 | 2020-04-10 | 东软医疗系统股份有限公司 | Training method and device of generative confrontation network, image enhancement method and equipment |
| US20210304364A1 (en) * | 2020-03-26 | 2021-09-30 | Bloomberg Finance L.P. | Method and system for removing noise in documents for image processing |
| WO2022041521A1 (en) * | 2020-08-31 | 2022-03-03 | 浙江大学 | Low-dose sinogram denoising and pet image reconstruction method based on teacher-student generators |
| CN113808042A (en) * | 2021-09-16 | 2021-12-17 | 沈阳工业大学 | SAR image denoising method based on wavelet transformation and generation countermeasure network |
Non-Patent Citations (2)
| Title |
|---|
| X. WANG,ET AL.: "A Self-Supervised Denoising Network for Satellite-Airborne-Ground Hyperspectral Imagery", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》 * |
| 冯旭斌: "基于深度学习的光学遥感图像去噪与超分辨率重建算法研究", 《CNKI博士论文》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115546653A (en) * | 2022-12-02 | 2022-12-30 | 耕宇牧星(北京)空间科技有限公司 | Remote sensing image denoising method based on depth enhancement network |
| CN116071265A (en) * | 2023-02-21 | 2023-05-05 | 哈尔滨工业大学 | Image denoising method based on space self-adaptive self-supervision learning |
| CN116071265B (en) * | 2023-02-21 | 2023-08-08 | 哈尔滨工业大学 | Image denoising method based on space self-adaptive self-supervision learning |
| CN116861167A (en) * | 2023-06-12 | 2023-10-10 | 河北工程大学 | FBG spectrum cyclic denoising method based on deep learning |
| CN116861167B (en) * | 2023-06-12 | 2023-12-15 | 河北工程大学 | FBG spectrum cyclic denoising method based on deep learning |
| CN117876692A (en) * | 2024-03-11 | 2024-04-12 | 中国石油大学(华东) | Feature weighted connection guided single-image remote sensing image denoising method |
| CN117876692B (en) * | 2024-03-11 | 2024-05-17 | 中国石油大学(华东) | Feature weighted connection guided single-image remote sensing image denoising method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114511473B (en) | 2022-07-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114511473B (en) | Hyperspectral remote sensing image denoising method based on unsupervised adaptive learning | |
| CN109035142B (en) | Satellite image super-resolution method combining countermeasure network with aerial image prior | |
| Vasamsetti et al. | Wavelet based perspective on variational enhancement technique for underwater imagery | |
| CN111028177A (en) | Edge-based deep learning image motion blur removing method | |
| Sonogashira et al. | High-resolution bathymetry by deep-learning-based image superresolution | |
| CN111008936B (en) | Multispectral image panchromatic sharpening method | |
| CN110648292A (en) | High-noise image denoising method based on deep convolutional network | |
| CN111179196B (en) | Multi-resolution depth network image highlight removing method based on divide-and-conquer | |
| CN112767267B (en) | Image defogging method based on simulation polarization fog-carrying scene data set | |
| CN111738954B (en) | Single-frame turbulence degradation image distortion removal method based on double-layer cavity U-Net model | |
| CN112330569A (en) | Model training method, text denoising method, device, equipment and storage medium | |
| CN110807428B (en) | Coal sample identification method, device, server and storage medium | |
| Feng et al. | Turbugan: An adversarial learning approach to spatially-varying multiframe blind deconvolution with applications to imaging through turbulence | |
| Hai et al. | Advanced retinexnet: a fully convolutional network for low-light image enhancement | |
| US20220067882A1 (en) | Image processing device, computer readable recording medium, and method of processing image | |
| Li et al. | Efficient burst raw denoising with variance stabilization and multi-frequency denoising network | |
| CN115760641B (en) | Remote sensing image cloud and fog removing method and equipment based on multiscale characteristic attention network | |
| Huang et al. | Attention-based for multiscale fusion underwater image enhancement | |
| CN115829870A (en) | Image denoising method based on variable scale filtering | |
| CN114758030A (en) | Underwater polarization imaging method integrating physical model and deep learning | |
| Su et al. | Deconvolution of defocused image with multivariate local polynomial regression and iterative wiener filtering in DWT domain | |
| CN113450267A (en) | Transfer learning method capable of rapidly acquiring multiple natural degradation image restoration models | |
| He et al. | An Unsupervised Dehazing Network with Hybrid Prior Constraints for Hyperspectral Image | |
| Sangeetha et al. | Performance analysis of exemplar based image inpainting algorithms for natural scene image completion | |
| CN114359103B (en) | Hyperspectral image defogging method, hyperspectral image defogging device, computer product and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |