CN113808042B - SAR image denoising method based on wavelet transformation and generation countermeasure network - Google Patents
SAR image denoising method based on wavelet transformation and generation countermeasure network Download PDFInfo
- Publication number
- CN113808042B CN113808042B CN202111084398.8A CN202111084398A CN113808042B CN 113808042 B CN113808042 B CN 113808042B CN 202111084398 A CN202111084398 A CN 202111084398A CN 113808042 B CN113808042 B CN 113808042B
- Authority
- CN
- China
- Prior art keywords
- image
- wavelet
- noise
- denoising
- generator
- 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.)
- Active
Links
- 230000009466 transformation Effects 0.000 title claims abstract description 28
- 238000000034 method Methods 0.000 title claims abstract description 27
- 238000012549 training Methods 0.000 claims abstract description 27
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 26
- 230000006870 function Effects 0.000 claims abstract description 19
- 230000002787 reinforcement Effects 0.000 claims abstract description 13
- 125000004122 cyclic group Chemical group 0.000 claims abstract description 6
- 238000004519 manufacturing process Methods 0.000 claims abstract description 3
- 238000010586 diagram Methods 0.000 claims description 14
- 230000000694 effects Effects 0.000 claims description 13
- 230000008569 process Effects 0.000 claims description 6
- 238000013507 mapping Methods 0.000 claims description 5
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 4
- 238000011176 pooling Methods 0.000 claims description 4
- 230000009467 reduction Effects 0.000 claims description 4
- 230000001105 regulatory effect Effects 0.000 claims description 4
- 238000004458 analytical method Methods 0.000 claims description 3
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 2
- 238000004364 calculation method Methods 0.000 claims description 2
- 238000005259 measurement Methods 0.000 claims description 2
- 238000010606 normalization Methods 0.000 claims description 2
- 238000007781 pre-processing Methods 0.000 claims description 2
- 238000012216 screening Methods 0.000 claims description 2
- 230000008901 benefit Effects 0.000 abstract description 2
- 230000009471 action Effects 0.000 description 6
- 238000011156 evaluation Methods 0.000 description 4
- 238000001914 filtration Methods 0.000 description 3
- 238000004445 quantitative analysis Methods 0.000 description 3
- 238000003384 imaging method Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000007704 transition Effects 0.000 description 2
- 230000008485 antagonism Effects 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/10—Image enhancement or restoration using non-spatial domain filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
- G06T2207/10044—Radar image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20048—Transform domain processing
- G06T2207/20064—Wavelet transform [DWT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a SAR image denoising method based on wavelet transformation and generation countermeasure network, which comprises the following steps: 1. manufacturing an SAR training sample image set; 2. performing wavelet transformation on an original image in the image set; 3. taking the wavelet transformed image as a training sample, and training by using a cyclic generation countermeasure network; 4. judging the denoising result of the trained network; 5. changing wavelet basis function to compare denoising results; 6. and (5) searching the decomposition scale of the optimal wavelet transformation by using reinforcement learning, and further improving the denoising result. The invention combines the advantages of retaining image details when the image pixels are operated by the countermeasure network by using the principle of automatic searching of wavelet decomposition scales, more pointedly removes the noise of high-frequency parts in different decomposition scales, and protects useful information in SAR images to a greater extent.
Description
Technical Field
The invention relates to the technical field of digital image processing, in particular to a SAR image denoising method based on wavelet transformation of automatic search scale and generation of an countermeasure network.
Background
Synthetic Aperture Radar (SAR) is a high resolution imaging radar that can obtain high resolution radar images resembling electrophotography under meteorological conditions of extremely low visibility. In the process of generating and transmitting SAR images, due to the influence of external environment and imaging equipment, certain random noise can be generated on the images, the quality of the images is greatly influenced, difficulties are brought to subsequent processing, the efficiency of subsequent feature extraction, recognition and classification and other works can be directly influenced by the good effect of denoising, and therefore the denoising of the SAR images is very important. The traditional image denoising method mainly adopts a filtering mode, directly processes pixel points of an original image, carries out convolution operation on the image and a set filter, adopts a classical algorithm including median filtering, self-adaptive filtering and the like, filters all pixels in the image, effectively filters noise, and simultaneously smoothes the image to damage edge details and textures of the image.
The image denoising method based on deep learning can autonomously learn noise characteristics due to strong learning capacity, realizes the mapping from the noisy picture to the noiseless picture, does not need to manually set denoising parameters, and is widely applied. In the existing image denoising method combining wavelet transformation and convolutional neural network, on the setting of wavelet decomposition scale, only one scale or two scales of decomposition is usually manually set, the setting only makes the method simpler, but the optimal denoising effect may not be achieved, the scale of wavelet decomposition directly affects the separation condition of the picture and noise contained in the picture, and based on the reason, the invention provides a method for automatically searching the wavelet transformation scale, combining the method with the generation of an countermeasure network, and automatically searching the optimal decomposition scale through reinforcement learning, thereby obtaining the optimal denoising effect.
Disclosure of Invention
The invention aims to provide a SAR image denoising method based on wavelet transformation and an antagonism network, which aims to solve the problem that the image cannot achieve the optimal denoising effect, so that the texture information and the building contour information in the image are blurred to generate information loss.
In order to achieve the above purpose, the present invention provides the following technical solutions: a method of denoising a SAR image based on wavelet transformation and generation of an countermeasure network, the method comprising the steps of: 1. manufacturing an SAR training sample image set; 2. performing wavelet transformation on an original image in the image set; 3. taking the wavelet transformed image as a training sample, and training by using a cyclic generation countermeasure network; 4. judging the denoising result of the trained network; 5. changing wavelet basis function to compare denoising results; 6. and (5) searching the decomposition scale of the optimal wavelet transformation by using reinforcement learning, and further improving the denoising result.
Compared with the prior art, the invention has the beneficial effects that:
the invention combines the advantages of retaining image details when the image pixels are operated by the countermeasure network by using the principle of automatic searching of wavelet decomposition scales, more pointedly removes the noise of high-frequency parts in different decomposition scales, and protects useful information in SAR images to a greater extent.
The evaluation of the denoising effect mainly comprises subjective qualitative evaluation and quantitative analysis of human vision, the quantitative analysis mainly comprises analysis of peak signal-to-noise ratio (PSNR) and Structural Similarity (SSIM) of a denoising image, and the denoising effect of the model is measured by the two indexes. The remote sensing image after denoising still has higher definition from the subjective view, so that the outline information of buildings, lanes, vehicles, ships and the like in the image is reserved while noise interference is removed to the greatest extent, noise components in each frequency component can be removed in a targeted manner, and useful information such as edge outlines and the like in the noise components is not influenced. In quantitative analysis, when standard deviation σ=20, the PSNR value is improved by 4.1% on average, and the SSIM value is improved by about 10.2% compared with the conventional method.
Drawings
Fig. 1 is a general flow chart of the present invention.
Fig. 2 is a schematic diagram of a two-dimensional wavelet transform.
Fig. 3 is a schematic diagram of a loop generation countermeasure network.
Fig. 4 is a training flow diagram of a loop generation countermeasure network.
Fig. 5 is a network structure diagram of a generator for cyclically generating a countermeasure network.
Fig. 6 is a diagram of a network structure of a discriminator for cyclically generating a countermeasure network.
FIG. 7 is a flow chart of a reinforcement learning search strategy.
FIG. 8 is a schematic diagram of the steps of the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Referring to fig. 1-8, the present invention provides a technical solution: a SAR image denoising method based on wavelet transformation and generation countermeasure network is mainly divided into two parts of training and searching, wherein the training process is shown in figure 1, the original noisy image is subjected to wavelet decomposition of different scales to obtain picture components of different sub-bands, then the picture components are respectively introduced into CycleGAN to obtain four components which are correspondingly output, and then the four components are subjected to wavelet inverse transformation and synthesized into a denoised image. The reinforcement learning part continuously changes the wavelet decomposition scale by reinforcement learning as shown in fig. 7, and finally searches out the optimal decomposition scale.
Step one: making SAR training sample image set
The training set selects a UC Merced Land-Use image data set, 1300 remote sensing images with the size of 256 multiplied by 256 are obtained through screening and used as a training sample set, and an OpenCV library function is used for adding Gaussian noise with the standard deviation of 20 into the training sample set to be used as a noise-containing sample set.
Step two: wavelet transforming original image in image set
As shown in fig. 2, for a two-dimensional image of size n×n, performing a wavelet transform of scale 1 first requires performing a wavelet transform of N one-dimensional data of length N in rows, and decomposing the image into two parts: the left side is a low frequency sub-image with the size of N multiplied by N/2, the right side is a high frequency sub-image with the size of N multiplied by N/2, then each sub-image is subjected to N/2 one-dimensional wavelet transformation with the length of N according to columns, and then the original image can be decomposed into 4 sub-images: LL1, HL1, LH1, HH1, wherein all three high frequency components contain more edge contours and noise information.
Step three: training with cyclic generation of countermeasure network by using wavelet transformed image as training sample
The loop generation countermeasure network training process is shown in fig. 4, wherein the Generator G and the Generator F are respectively a noise-free image and a noise-containing image Generator, and the identifier G and the identifier F are respectively a noise-free image and a noise-containing image Discriminator, which aim to judge whether the input image is a real picture or is generated by the Generator; the parameters of the network are mainly regulated by the counter propagation of two types of losses, the first type is the counter Loss (universal Loss), as shown by dark arrows in fig. 4, a real noisy image is changed into a generated noiseless image after passing through a generator G, then the noiseless image is judged by a discriminator G, if the real noisy image is judged to be a generated false picture by the discriminator, the parameters of the generator G need to be regulated, and the generator F is similar; the second type is a loss of cyclical consistency (CycleConsistency Loss), as shown by the light arrows in fig. 4, after the actual noisy image passes through the generator G and is converted into a generated noiseless image, the noiseless image is restored to the noisy image after the image is fed into the generator F, and the image secondarily generated by the two generators should be similar to the original actual image, so that the generator can only operate on the noisy part of the image and does not affect other parts of the image, and therefore, the parameters of the generators G and F are reversely adjusted by the similarity between the two images;
the generator structure adopts a structure similar to a U-net network, the first layers in the network adopt convolution layers to extract the picture characteristic information, the downsampling is adopted to reduce the size of a characteristic diagram, the later layers adopt inverse convolution to perform upsampling, the processed characteristic diagram is restored to the original diagram size and output, nine residual blocks are added between the downsampling and the upsampling, and a network model is shown in figure 5; filling a 7×7 convolution kernel with a mirror boundary in a first convolution block, so that the size of the feature map after convolution is kept unchanged, and then implementing downsampling by two convolution blocks, wherein each convolution block consists of a convolution layer with the size of 3×3, a single-channel internal normalization layer and a ReLU activation function layer; after that, through nine repeated residual blocks, the part is a main body of the generator, and mapping transformation is carried out on the feature map which is subjected to the pre-processing in the area; the processed feature map is up-sampled by two deconvolution blocks, and the image is amplified; finally, carrying out boundary filling through an output layer and carrying out convolution kernel operation with 9 multiplied by 9 to restore the picture to the original size and output; the judging structure is composed of five layers of convolutions and the output of the last pooling layer, and the judging device only needs to complete the two classification tasks generated by the judging device for judging whether the picture is generated or not, so that the number of channels is reduced to 1 only through five 4×4 convolutions, and the output of the single channel with the size of 1×1 is finally obtained through average pooling.
Step four: judging denoising result of trained network
The denoising effect of the model is measured by adopting two indexes of peak signal-to-noise ratio (PSNR) and Structural Similarity (SSIM) of the image; the PSNR is an important measurement index in the image noise reduction problem, and is defined by the mean square error of the image, wherein the larger the numerical value is, the better the noise reduction degree is represented, and the calculation is carried out according to the following formula;
W×H is the resolution of the image, I and I 0 Representing the denoised image and the initial image; SSIM is used for evaluating the similarity degree of two images from three aspects of brightness, contrast and structure, the value range is between 0 and 1, the larger the value is, the closer the image is to the original image compared with the original image, the image information is complete, and the denoising effect is good, as shown in formulas 3 to 6;
SSIM(I,I 0 )=[l(I,I 0 )] α [c(I,I 0 )] β [s(I,I 0 )] γ 3)
wherein l (I, I) 0 ) Is a brightness comparison, c (I, I 0 ) Is contrast ratio, s (I, I 0 ) Is a structural comparison, mu I Andrespectively represent I and I 0 Mean value of σ I And->Respectively represent I and I 0 Standard deviation of>Represents I and I 0 Covariance of c 1 ,c 2 ,c 3 Each constant, α=β=γ=1 is set.
Step five: altering wavelet basis function versus denoising results
In the wavelet transformation process, the final denoising effect can be influenced by using different wavelet basis functions; there are therefore four common wavelet basis functions to compare: haar, daubechies, symlet and coidlet final denoising results; the four wavelet basis functions are used for decomposition in the first step, and the haar wavelet basis functions are obtained through analysis and comparison to be optimal after the decomposition in the first step is sequentially carried out to the fourth step.
Step six: searching for the decomposition scale of the optimal wavelet transform using reinforcement learning to further enhance the denoising result
As shown in fig. 7, the reinforcement learning controller is configured by a cyclic neural network, different wavelet decomposition scales are controlled and selected, then a corresponding denoising result diagram is obtained through network training of the first to fourth steps, the scales of wavelet decomposition are corrected through obtained PSNR and SSIM judgment results, so that iteration is continued, and finally the optimal decomposition scales are converged.
Working principle:
the method comprises a wavelet transformation part, a training part formed by a generating countermeasure network part and an reinforcement learning automatic search scale part, wherein the image is subjected to wavelet change and decomposed into high-frequency and low-frequency different sub-bands through reinforcement learning strategies, then wavelet coefficients needing to be kept in the generation network training learning are used, finally the frequency domain image is converted back to an original space domain through wavelet inversion to obtain a denoised picture, the denoised effect is evaluated to obtain a return, and reinforcement learning is guided to perform search of different scales.
The purpose of the wavelet transformation part is to transform the original SAR image containing noise from the spatial domain to the frequency domain, so as to separate the whole useful information in the picture from the high-frequency noise components existing in the picture, and the decomposition adopts different basis functions to have a certain influence on the result.
The function of the loop generation countermeasure network part is to reject noise components from different components obtained after wavelet transformation. Compared with the original GAN network, the cyclic generation countermeasure network (cyclGAN) is additionally provided with a generator and a discriminator, the generator aims at cheating the discriminator to generate false samples, and the discriminator distinguishes whether the input is a real sample or is generated by the generator as accurately as possible, and the two are in continuous conflict opposition, so that the performance is improved together. The CycleGAN model as shown in fig. 3, the training purpose is to learn the mapping between two domains, X and Y, G and F being the generator of the Y domain and the generator of the X domain, respectively, dx and Dy being the discriminators of the X and Y domains, respectively. The loss function of the network consists of two parts: counter Loss (universal Loss) and loop consistency Loss (Cycle Consistency Loss). Parameters of the generator are updated in the network model using the countermeasures against losses, which should be for the generator G to transition the image from the X-domain to the Y-domain:
DY is the arbiter of the Y-domain image, G (X) is the generator of the Y-domain image generated from the X-domain image, similarly for generator F to transition the image from Y-domain to X-domain, the countermeasures against loss should be:
DX is a discriminator of the X-domain image, and F (Y) is an X-domain image generated by the generator from the Y-domain image. The mere use of countermeasures cannot guarantee the quality of the generated pictures, and the mere use of countermeasures may cause the network to map all the images in the X domain to one picture in the Y domain, and for applications in denoising the pictures, the useful information in the images is lost entirely. Therefore, a cycle consistency loss is introduced, and for each sample image X of the domain X, x→g (X) →f (G (X))≡x is required, which is called forward cycle consistency of the network; similarly for image Y within domain Y, generators G and F should also satisfy reverse loop consistency, i.e., y→f (Y) →g (F (Y))≡y, the loop consistency loss of the network is expressed as follows:
the reinforcement learning is used as a learning system to learn in a trial-and-error mode, acquires current state information s from an external environment, takes heuristic action u on the environment, and acquires evaluation r of the action and a new environment state of environment feedback. If an action u of the agent causes a positive prize to the environment, the agent's tendency to produce this action later will be enhanced; conversely, the tendency of the agent to produce this action will diminish. In the repeated interaction of the control behavior of the learning system with the state and evaluation of the environmental feedback, the mapping strategy from the state to the action is continuously modified in a learning mode so as to obtain the optimal waiting result.
Claims (5)
1. A SAR image denoising method based on wavelet transformation and generation countermeasure network is characterized in that: the method comprises the following steps:
1. manufacturing an SAR training sample image set;
2. performing wavelet transformation on an original image in the image set;
3. taking the wavelet transformed image as a training sample, and training by using a cyclic generation countermeasure network;
4. judging the denoising result of the trained network;
5. changing wavelet basis function to compare denoising results;
6. searching the decomposition scale of the optimal wavelet transformation by using reinforcement learning, and further improving the denoising result;
the third step is as follows:
the parameters of the network are regulated by the counter propagation of two types of losses, the first type is countering the losses, a real noise-containing image is changed into a generated noise-free image after passing through a generator G, then the noise-free image is judged by a discriminator G, if the false image is judged to be generated by the discriminator, the parameters of the generator G need to be regulated, and the generator F is the same; the second type is that the circulation consistency is lost, after the real noise-containing image is changed into the generated noise-free image after passing through the generator G, the noise-containing image is restored into the noise-containing image after passing through the generator F, and the image which is secondarily generated by the two generators is approximate to the original real image, so that the generator can only operate the noise part in the image without affecting other parts of the image, and the parameters of the generator G and the generator F are reversely adjusted by using the similarity between the two parts;
the generator structure adopts a structure similar to a U-net network, the first layers in the network adopt convolution layers to extract the picture characteristic information, the size of a characteristic diagram is reduced by adopting downsampling, the later layers adopt inverse convolution to perform upsampling, the processed characteristic diagram is restored to the original diagram size and output, and nine residual blocks are added between the downsampling and the upsampling; filling a 7×7 convolution kernel with a mirror boundary in a first convolution block, so that the size of the feature map after convolution is kept unchanged, and then implementing downsampling by two convolution blocks, wherein each convolution block consists of a convolution layer with the size of 3×3, a single-channel internal normalization layer and a ReLU activation function layer; after that, through nine repeated residual blocks, the part is a main body of the generator, and mapping transformation is carried out on the feature map which is subjected to the pre-processing in the area; the processed feature map is up-sampled by two deconvolution blocks, and the image is amplified; finally, carrying out boundary filling through an output layer and carrying out convolution kernel operation with 9 multiplied by 9 to restore the picture to the original size and output; the judging structure is composed of five layers of convolutions and the output of the last pooling layer, and the judging device only needs to complete the two classification tasks generated by the judging device for judging whether the picture is generated or not, so that the number of channels is reduced to 1 only through five 4×4 convolutions, and the output of the single channel with the size of 1×1 is finally obtained through average pooling;
the fourth step is specifically as follows:
the denoising effect of the model is measured by adopting two indexes of peak signal-to-noise ratio (PSNR) and Structural Similarity (SSIM) of the image; the PSNR is an important measurement index in the image noise reduction problem, and is defined by the mean square error of the image, wherein the larger the numerical value is, the better the noise reduction degree is represented, and the calculation is carried out according to the following formula;
W×H is the resolution of the image, I and I 0 Representing the denoised image and the initial image;
SSIM is used for evaluating the similarity degree of two images from three aspects of brightness, contrast and structure, the value range is between 0 and 1, the larger the value is, the closer the image is to the original image compared with the original image, the image information is complete, and the denoising effect is good, as shown in formulas 3 to 6;
SSIM(I,I 0 )=[l(I,I 0 )] a [c(I,I 0 )] β [s(I,I 0 )] γ 3)
wherein l (I, I) 0 ) Is a brightness comparison, c (I, I 0 ) Is contrast ratio, s (I, I 0 ) Is a structural comparison, mu I Andrespectively represent I and I 0 Mean value of σ I And->Respectively represent I and I 0 Standard deviation of>Represents I and I 0 Covariance of c 1 ,c 2 ,c 3 Each constant, α=β=γ=1 is set.
2. The SAR image denoising method based on wavelet transform and generation countermeasure network according to claim 1, wherein: the first step is as follows:
the training set selects a UC Merced Land-Use image data set, 1300 remote sensing images with the size of 256 multiplied by 256 are obtained through screening to serve as a training sample set, and an OpenCV library function is used for adding Gaussian noise with the standard deviation of 20 into the training sample set to serve as a noise-containing sample set.
3. The SAR image denoising method based on wavelet transform and generation countermeasure network according to claim 2, wherein: the second step is specifically as follows:
performing wavelet transformation with a scale of 1 on a two-dimensional image with a size of N multiplied by N; first, wavelet transformation of N one-dimensional data with length N is needed to be performed according to the line, and an image is decomposed into two parts: the left side is a low frequency sub-image with the size of N multiplied by N/2, the right side is a high frequency sub-image with the size of N multiplied by N/2, then each sub-image is subjected to N/2 one-dimensional wavelet transformation with the length of N according to columns, and then the original image can be decomposed into 4 sub-images: LL1, HL1, LH1, HH1, wherein all three high frequency components contain more edge contours and noise information.
4. The SAR image denoising method based on wavelet transform and generation countermeasure network according to claim 1, wherein: the fifth step is specifically as follows:
in the wavelet transformation process, the final denoising effect is affected by using different wavelet basis functions; there are therefore four common wavelet basis functions to compare: haar, daubechies, symlet and coidlet final denoising results; the four wavelet basis functions are used for decomposition in the first step, and the haar wavelet basis functions are obtained through analysis and comparison to be optimal after the decomposition in the first step is sequentially carried out to the fourth step.
5. The SAR image denoising method based on wavelet transform and generation countermeasure network according to claim 1, wherein: the sixth step is specifically as follows:
the reinforced learning controller is composed of a circulating neural network, different wavelet decomposition scales are controlled and selected by the circulating neural network, then a corresponding denoising result diagram is obtained through network training of the first to fourth steps, the scales of wavelet decomposition are corrected through obtained PSNR and SSIM judgment results, iteration is continued, and finally the optimal decomposition scales are converged.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111084398.8A CN113808042B (en) | 2021-09-16 | 2021-09-16 | SAR image denoising method based on wavelet transformation and generation countermeasure network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111084398.8A CN113808042B (en) | 2021-09-16 | 2021-09-16 | SAR image denoising method based on wavelet transformation and generation countermeasure network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113808042A CN113808042A (en) | 2021-12-17 |
CN113808042B true CN113808042B (en) | 2024-04-05 |
Family
ID=78895471
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111084398.8A Active CN113808042B (en) | 2021-09-16 | 2021-09-16 | SAR image denoising method based on wavelet transformation and generation countermeasure network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113808042B (en) |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114511473B (en) * | 2022-04-19 | 2022-07-05 | 武汉大学 | Hyperspectral remote sensing image denoising method based on unsupervised adaptive learning |
CN115082296B (en) * | 2022-07-11 | 2024-05-14 | 杭州师范大学 | Image generation method based on wavelet domain image generation frame |
CN115310488A (en) * | 2022-08-16 | 2022-11-08 | 哈尔滨工业大学 | Seismic oscillation recording filtering method based on generating type antagonistic neural network |
CN117076875B (en) * | 2023-10-18 | 2024-01-26 | 中核武汉核电运行技术股份有限公司 | Denoising method for nuclear signal under complex noise background |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101482617A (en) * | 2009-01-16 | 2009-07-15 | 西安电子科技大学 | Synthetic aperture radar image denoising method based on non-down sampling profile wave |
CN110517195A (en) * | 2019-07-26 | 2019-11-29 | 西安电子科技大学 | Unsupervised SAR image denoising method |
CN112037151A (en) * | 2020-09-08 | 2020-12-04 | 天元大数据信用管理有限公司 | Image denoising method based on wavelet analysis |
-
2021
- 2021-09-16 CN CN202111084398.8A patent/CN113808042B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101482617A (en) * | 2009-01-16 | 2009-07-15 | 西安电子科技大学 | Synthetic aperture radar image denoising method based on non-down sampling profile wave |
CN110517195A (en) * | 2019-07-26 | 2019-11-29 | 西安电子科技大学 | Unsupervised SAR image denoising method |
CN112037151A (en) * | 2020-09-08 | 2020-12-04 | 天元大数据信用管理有限公司 | Image denoising method based on wavelet analysis |
Also Published As
Publication number | Publication date |
---|---|
CN113808042A (en) | 2021-12-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113808042B (en) | SAR image denoising method based on wavelet transformation and generation countermeasure network | |
CN109360156B (en) | Single image rain removing method based on image block generation countermeasure network | |
CN111242862B (en) | Multi-scale fusion parallel dense residual convolution neural network image denoising method | |
Zhang et al. | Joint image denoising using adaptive principal component analysis and self-similarity | |
CN108447041B (en) | Multi-source image fusion method based on reinforcement learning | |
CN109035189B (en) | Infrared and weak visible light image fusion method based on Cauchy fuzzy function | |
CN113837974B (en) | NSST domain power equipment infrared image enhancement method based on improved BEEPS filtering algorithm | |
CN104657951A (en) | Multiplicative noise removal method for image | |
Min et al. | Blind deblurring via a novel recursive deep CNN improved by wavelet transform | |
Liu et al. | True wide convolutional neural network for image denoising | |
CN116596792B (en) | Inland river foggy scene recovery method, system and equipment for intelligent ship | |
CN114581318B (en) | Low-illumination image enhancement method and system | |
Xu et al. | Remote sensing image denoising using patch grouping-based nonlocal means algorithm | |
CN116309178A (en) | Visible light image denoising method based on self-adaptive attention mechanism network | |
CN117726537A (en) | SAR image denoising network method and system for self-adaptive multi-scale feature fusion AMFFD-Net | |
CN115082336A (en) | SAR image speckle suppression method based on machine learning | |
Zhao et al. | Learning a deep convolutional network for subband image denoising | |
CN115689958A (en) | Synthetic radar image denoising method based on deep learning | |
CN115937302A (en) | Hyperspectral image sub-pixel positioning method combined with edge preservation | |
CN115082296A (en) | Image generation method based on wavelet domain image generation framework | |
CN113177904A (en) | Image fusion method and system | |
CN113379641A (en) | Single image rain removing method and system based on self-coding convolutional neural network | |
CN106651804B (en) | Vehicle-mounted image denoising method and system based on adaptive diffusion filtering | |
Chen et al. | GADO-Net: an improved AOD-Net single image dehazing algorithm | |
Tang et al. | How SAR Image Denoise Affects the Performance of DCNN-Based Target Recognition Method |
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 |