CN110674930A - SAR image denoising method based on learning down-sampling and jump connection network - Google Patents
SAR image denoising method based on learning down-sampling and jump connection network Download PDFInfo
- Publication number
- CN110674930A CN110674930A CN201910927416.0A CN201910927416A CN110674930A CN 110674930 A CN110674930 A CN 110674930A CN 201910927416 A CN201910927416 A CN 201910927416A CN 110674930 A CN110674930 A CN 110674930A
- Authority
- CN
- China
- Prior art keywords
- image
- sar
- learning
- downsampling
- method based
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 44
- 238000005070 sampling Methods 0.000 title claims description 11
- 238000012549 training Methods 0.000 claims description 20
- 238000013527 convolutional neural network Methods 0.000 claims description 19
- 230000006870 function Effects 0.000 claims description 16
- 238000003062 neural network model Methods 0.000 claims description 9
- 230000002441 reversible effect Effects 0.000 claims description 7
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 4
- 230000008034 disappearance Effects 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 101001013832 Homo sapiens Mitochondrial peptide methionine sulfoxide reductase Proteins 0.000 claims description 3
- 102100031767 Mitochondrial peptide methionine sulfoxide reductase Human genes 0.000 claims description 3
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 abstract description 5
- 238000002474 experimental method Methods 0.000 abstract description 3
- 208000001840 Dandruff Diseases 0.000 abstract 1
- 238000013507 mapping Methods 0.000 abstract 1
- 230000001629 suppression Effects 0.000 abstract 1
- 230000000694 effects Effects 0.000 description 7
- 230000009466 transformation Effects 0.000 description 5
- 238000004458 analytical method Methods 0.000 description 4
- 238000001914 filtration Methods 0.000 description 4
- 230000001427 coherent effect Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 238000001514 detection method Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 230000005855 radiation Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001066 destructive effect Effects 0.000 description 1
- 238000011438 discrete method Methods 0.000 description 1
- 238000002592 echocardiography Methods 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 239000002245 particle Substances 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
Images
Classifications
-
- G06T5/70—
-
- 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
-
- 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
-
- 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
- G06T2207/20081—Training; Learning
Abstract
The invention discloses a SAR image denoising method based on learning downsampling and jump connection network, which realizes nonlinear end-to-end mapping between a noise image and a clean SAR image by utilizing the learning downsampling and jump connection network (SAR-DSCN). A jump connection network is added into a denoising model to keep the details of an image and reduce the problem of vanishing gradients. The use of downsampling also allows the receive domain to be effectively extended. A large number of experiments on SAR images show that the method has better performance and higher speed than the most advanced speckle suppression method at present. The result shows that the effectiveness and high efficiency of the SAR-DSCN make the application of the SAR-DSCN in the SAR image anti-dandruff processing have certain attraction.
Description
Technical Field
The invention relates to the technical field of SAR image denoising, in particular to an SAR image denoising method based on learning downsampling and jump connection network.
Background
Synthetic Aperture Radar (SAR) is an active coherent microwave radar that produces images with high spatial resolution. The method has the characteristics of all-weather imaging, day and night imaging, high resolution and the like, has important application value in the field of remote sensing, and plays an important role in the fields of military and civil use. However, under coherent radiation, the original image has a backscattering coefficient and the particle noise is uniform, i.e., speckle noise, within the same area. It is caused by constructive and destructive interference of coherent echoes scattered by the smal reflectors within each resolution cell. The presence of speckle noise in SAR images often presents difficulties to the processing and interpretation of computer vision systems and human interpreters. Therefore, removing speckle noise is a key task in preprocessing, and is an indispensable part of segmentation, detection and classification in subsequent image processing.
The existing solutions for denoising SAR images mainly include: the method is based on a multi-scale geometric analysis method and adopts the traditional methods of spatial filtering and frequency filtering, and the image denoising method based on wavelet transformation. However, the above-mentioned several image denoising schemes have their respective drawbacks:
for the spatial domain denoising algorithm, the method can better suppress noise and can save texture information, edges, linear characteristics and point target response. However, due to the nature of local processing, the spatial linear filtering method often cannot preserve edges and details completely. It suffers from the disadvantages of 1) the inability to maintain an average value, particularly for equivalent SAR images, with a small horizon (ENL); 2) specific targets such as strongly reflected points, small surface features, etc. are easily blurred or erased; 3) speckle noise in dark scenes is not removed.
For the frequency domain denoising algorithm, the edge and part of high-frequency texture information are often blurred due to the ringing phenomenon in the frequency domain image denoising process, and the information is lost when being mapped through frequency domain transformation, so that part of high-frequency information is lost in the image.
The wavelet transform-based image denoising method, which eliminates noise by filtering wavelet coefficients of a transform domain assuming that the noise is mainly present in high-frequency wavelet components, is very successful in reducing additive white gaussian noise. In order to apply wavelet transform to SAR denoising, a logarithmic transform is generally used to convert speckle noise into gaussian noise. However, the logarithm operation often distorts the radiation characteristic of the SAR image after denoising.
For the multi-scale geometric analysis method, the profile wave and the shear wave are most widely applied in image denoising. The profile wave not only has enough directivity, but also maintains the multi-scale characteristic and the time-frequency local characteristic of the wavelet. However, the contourlet transformation is based on a discrete method, and does not conform to the multi-resolution analysis (MRA) theory, and various difficulties exist in mathematical theory analysis. Shear waves have the advantages of a profile line and overcome the disadvantages thereof. The Shearlet transformation not only accords with MRA in theory, but also can perform sparse representation on the image to generate optimal approximation, so that the Shearlet transformation can be used for image processing tasks such as edge extraction and target detection. The Shearlet transform is easier to implement while achieving flexible directional selectivity. However, it also has some disadvantages, such as poor translational robustness and edge pseudo-gibbs distortion.
Disclosure of Invention
The invention aims to solve the problems that: the SAR image denoising method based on the learning downsampling and jump connection network is provided, the performance is excellent under the environments with different speckle levels, the algorithm speed is increased on the premise of not sacrificing the speckle removing performance, and no visual artifact is introduced when the balance between noise reduction and detail storage is controlled.
The technical scheme provided by the invention for solving the problems is as follows: a SAR image denoising method based on learning down-sampling and jump connection network comprises the following steps,
s1: building an integral depth convolution neural network model consisting of reversible down-sampling, a noise estimation part and up-sampling, and adding the final output of the neural network and a source image to form a residual learning layer;
s2: selecting a training set, and setting the learning rate, the attenuation rate and the training times of the deep convolutional neural network model in the step S1;
s3: setting a loss function according to the hyper-parameters set in the step S2 and the structure of the deep convolutional neural network in the step S1;
s4: continuously minimizing the loss function set in the step S3 by adopting a random gradient descent algorithm to obtain a new deep convolution neural network model;
s5: and (4) inputting the SAR noise image into the new deep convolutional neural network model finally obtained in the step S4, and outputting the denoised SAR image.
Preferably, the reversible downsampling in the overall depth convolutional neural network model built in step S1 reconstructs the input source image into four sub-images with a size of one fourth of the source image, and the four sub-images are used as the input of the CNN, so that the receiving domain can be effectively expanded, and the algorithm rate is improved.
Preferably, the noise estimation portion in the overall deep convolutional neural network model constructed in step S1 is composed of 12 convolutional layers and batch normalization and ReLU activation functions, and the feature information of the previous layer can be transferred to the next layer by using a jump connection structure, so that image details can be maintained, and the problem of gradient disappearance in the deep network can be avoided or reduced.
Preferably, the estimated speckle noise is subtracted from the output sub-image to obtain a denoised sub-image.
Preferably, the denoised sub-image is up-sampled to obtain a denoised image.
Preferably, the training set in step S2 is obtained by adding speckle of different levels to a clean data set, so as to obtain a training set of SAR images with different noise levels, and the included noise image is divided into noise image blocks of 40 × 40 with a step size of 128.
Preferably, the loss function in step S3 is an euclidean distance function:
in the formula, theta represents a network structure parameter, N is the total number of samples, and Y isiAnd XiThe speckle images and the original images of the images in the training set.
Preferably, the weight initialization of the network model in step S4 is generated by using MSRA algorithm, and the random gradient descent algorithm uses Adam optimization algorithm modified by the random gradient descent algorithm.
Compared with the prior art, the invention has the advantages that:
1. reversible downsampling in the built integral depth convolution neural network model enables an input source image to be reconstructed into four sub-images with the size being one fourth of the size of the source image, the sub-images serve as input of the CNN, a receiving domain can be effectively expanded, and therefore algorithm speed is improved.
2. The noise estimation part in the overall depth convolutional neural network model built in step S1 is composed of 12 convolutional layers (and batch normalization and ReLU activation functions), and the feature information of the previous layer can be transferred to the next layer by using a hopping connection structure, so that the image details can be maintained, and the problem of gradient disappearance in the depth network can be avoided or reduced.
3. Residual error learning is added, a better learning effect can be obtained through a multi-layer network, and training loss is reduced rapidly. And residual learning can effectively transmit different levels of feature information between indirectly connected layers without attenuation. Residual learning can translate the gradient descent process into a much smoother super-surface loss to the filter parameters. Thus, finding an assignment that is close to the optimum of the network parameters becomes faster and easier, allowing us to add more trainable layers to the network and improve its performance. By utilizing the learning process of the residual error unit, the original multiplicative speckle noise is more easily approximated through deeper and inherent linear feature extraction and expression, and the distance difference between the optical image and the SAR image can be better weakened.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the invention and not to limit the invention.
Fig. 1 is a flow chart of an SAR image denoising method based on learning downsampling and hopping connection networks according to the present invention.
FIG. 2 shows a jump connection structure
FIG. 3 is a network structure for image denoising.
Fig. 4 is a de-noising effect map of an airplane scene with L-4 for different methods (a) an original image (b) a speckle image (c) a front-filter (d) Shearlet (e) SAR-DM3D, (f) SAR-DRN (g) SAR-DSCN;
fig. 5 is a denoising effect map of a building scene in L-4 case for different methods (a) original image (b) speckle image (c) front-filter (d) Shearlet (e) SAR-DM3D, (f) SAR-DRN. (g) SAR-DSCN;
fig. 6 is a denoising effect map of a stadium scene with different methods in L-4 (a) an original image (b) a speckle image (c) front-filter (d) Shearlet (e) SAR-DM3D, (f) SAR-DRN. (g) SAR-DSCN.
Detailed Description
The following detailed description of the embodiments of the present invention will be provided with reference to the accompanying drawings and examples, so that how to implement the embodiments of the present invention by using technical means to solve the technical problems and achieve the technical effects can be fully understood and implemented.
As shown in fig. 1, a method for denoising an SAR image based on a learning downsampling and hopping connection network includes the following steps:
s1: building an integral depth convolution neural network model consisting of reversible down-sampling, a noise estimation part and up-sampling, and adding the final output of the neural network and a source image to form a residual learning layer;
reversible downsampling in the built integral depth convolution neural network model enables an input source image to be reconstructed into four sub-images with the size being one fourth of the size of the source image, the sub-images serve as input of the CNN, a receiving domain can be effectively expanded, and therefore algorithm speed is improved. The noise estimation section consists of 12 convolutional layers (and batch normalization and ReLU activation functions), more specifically, the first convolutional layer uses "Conv + ReLU", the middle layer uses "Conv + BN + ReLU", and the last convolutional layer uses "Conv". The padding is zero to ensure that the output of each layer is the same as the size of the input image, and meanwhile, the feature information of the previous layer can be transferred to the next layer by using a jump connection structure (as shown in fig. 2), so that the image details can be maintained, and the problem of gradient disappearance in a depth network can be avoided or reduced. And restoring the denoised sub-image by utilizing upsampling.
S2: selecting a training set, and setting the learning rate, the attenuation rate and the training times of the deep convolutional neural network model in the step S1;
in order to train the SAR-DSCN model, an input-output pair is prepared
The training data set of (1). Where Y is obtained by adding speckle to the latent image X (as shown in fig. 3). Considering that the SAR image without the speckles is difficult to obtain, UC Mercded land use dataset is used as a training data set for simulating SAR image denoising. It contains 21 scene classes, each containing 100 pictures. All images were adjusted to 256x 256. To train the proposed SAR-DSCN, we selected 400 images from this dataset and set the size of each image block to 40x 40. 193,664 image blocks are then cropped for training SAR-CNN, with a minimum batch size of 128, for parallel computation. Setting the depth of the image to 1; the learning rate was set to 0.0001, and the decay rate at each training was set to 0.0001; the test is carried out every 500 times of training, the effect of the current model and the corrected parameter value are observed, and the total number of iterative training is 1000000.
S3: setting a loss function according to the hyper-parameters set in the step S2 and the structure of the deep convolutional neural network in the step S1;
the loss function is the euclidean distance function:
in the formula, Θ represents a network structure parameter, N is a total number of samples, and Yi and Xi are a speckle image and an original image of the images in the training set.
S4: continuously minimizing the loss function set in the step S3 by using a random gradient descent algorithm to obtain a new deep convolution neural network model;
in the embodiment, Adam is selected as an algorithm for optimizing the network structure, wherein an Adam optimization method is an improved version of a gradient descent algorithm, and variable learning rate and attenuation are introduced. This is done by iterating once per time step and calculating the average gradient and the square root decay of the average gradient. Therefore, the surface gradient descent method can be converged to the defect of the local optimal solution to a certain extent. The MSRA is adopted in the weight initialization algorithm, so that the problem of slow convergence caused by different sizes of convolution kernels is solved. The generated weight value is random enough, so that the solution closest to the global optimum can be converged in a certain range.
S5: and inputting the noisy image into the model finally obtained in the step S4, and outputting the denoised image.
In order to test the performance of the proposed model, three categories of airplane, building and stadium are taken as simulation images respectively. In a real SAR image denoising experiment, two images, namely a classic Frost SAR image and a Snowberg SAR image, which are commonly used in real SAR data image denoising are used. The experiments were performed at four different noise levels 1, 2, 4, 8. The peak signal-to-noise ratio (PSNR) and the Structural Similarity Index (SSIM) are used as experimental evaluation indexes. By comparing the SAR-dscn method with Frost-Filter, SAR-BM3D, shear wave SAR image denoising and SAR-DRN. And best performance is indicated in bold.
As can be seen from table one, table two and table three, the SAR-DSCN model obtains 11 of the 12 best PSNR results and 8 of the 12 best SSIM results at four noise levels, and the overall effect of the SAR-DSCN method proposed herein is superior to the other four methods. From Table four we can see that the average PSNR value of this method is about 0.87dB/1.03dB/1.1dB/2.06dB higher than that of the SAR-CNN method, and from Table five we can see that the average value SSIM of this method is about 0.02/0.036/0.039/0.041 at different noise levels compared to that of the SAR-CNN method. As can be seen from fig. 4, 5 and 6, the method of the present invention has better performance than the existing methods in both quantitative and visual evaluation, especially under strong speckle noise.
Table one results of psnr (db) and SSIM for aircraft at L ═ 1, 2, 4, and 8
Table two results of psnr (db) and SSIM for buildings when L ═ 1, 2, 4, and 8
Table iii results of psnr (db) and SSIM for stadiums at L ═ 1, 2, 4, and 8
Table four average psnr (db) results for three different scenarios with different methods when L is 1, 2, 4, 8
Table five average SSIM results for three different scenarios with different methods when L ═ 1, 2, 4, and 8
The above-mentioned embodiments are merely preferred embodiments of the present invention, and the scope of the present invention is not limited thereto, so that variations based on the shape and principle of the present invention should be covered within the scope of the present invention.
The foregoing is merely illustrative of the preferred embodiments of the present invention and is not to be construed as limiting the claims. The present invention is not limited to the above embodiments, and the specific structure thereof is allowed to vary. All changes which come within the scope of the invention as defined by the independent claims are intended to be embraced therein.
Claims (8)
1. A SAR image denoising method based on learning down-sampling and jump connection network is characterized in that: the method comprises the following steps of,
s1: building an integral depth convolution neural network model consisting of reversible down-sampling, a noise estimation part and up-sampling, and adding the final output of the neural network and a source image to form a residual learning layer;
s2: selecting a training set, and setting the learning rate, the attenuation rate and the training times of the deep convolutional neural network model in the step S1;
s3: setting a loss function according to the hyper-parameters set in the step S2 and the structure of the deep convolutional neural network in the step S1;
s4: continuously minimizing the loss function set in the step S3 by adopting a random gradient descent algorithm to obtain a new deep convolution neural network model;
s5: and (4) inputting the SAR noise image into the new deep convolutional neural network model finally obtained in the step S4, and outputting the denoised SAR image.
2. The SAR image denoising method based on learning downsampling and hopping connection network as claimed in claim 1, wherein: reversible downsampling in the overall depth convolution neural network model built in the step S1 enables the input source image to be reconstructed into four sub-images with the size being one fourth of the source image, and the sub-images serve as input of the CNN, so that a receiving domain can be effectively expanded, and the algorithm speed is improved.
3. The SAR image denoising method based on learning downsampling and hopping connection network as claimed in claim 1, wherein: the noise estimation part in the overall depth convolutional neural network model built in the step S1 is composed of 12 convolutional layers and batch normalization and ReLU activation functions, and the feature information of the previous layer can be transferred to the next layer by using a jump connection structure, so that the image details can be maintained, and the problem of gradient disappearance in the depth network can be avoided or reduced.
4. The SAR image denoising method based on learning downsampling and hopping connection network as claimed in claim 2, characterized in that: and subtracting the estimated speckle noise from the output sub-image to obtain the denoised sub-image.
5. The SAR image denoising method based on learning downsampling and hopping connection network as claimed in claim 4, wherein: and the denoised sub-image is subjected to up-sampling to obtain a denoised image.
6. The SAR image denoising method based on learning downsampling and hopping connection network as claimed in claim 1, wherein: the training set in step S2 is obtained by adding speckles with different levels to a clean data set, so as to obtain a training set of SAR images with different noise levels, and the included noise image is divided into noise image blocks of 40 × 40 with a step size of 128.
7. The SAR image denoising method based on learning downsampling and hopping connection network as claimed in claim 1, wherein: the loss function in step S3 is an euclidean distance function:
in the formula, theta represents a network structure parameter, N is the total number of samples, and Y isiAnd XiThe speckle images and the original images of the images in the training set.
8. The SAR image denoising method based on learning downsampling and hopping connection network as claimed in claim 1, wherein: the weight initialization of the network model in the step S4 is generated by using the MSRA algorithm, and the random gradient descent algorithm uses its improved Adam optimization algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910927416.0A CN110674930A (en) | 2019-09-27 | 2019-09-27 | SAR image denoising method based on learning down-sampling and jump connection network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910927416.0A CN110674930A (en) | 2019-09-27 | 2019-09-27 | SAR image denoising method based on learning down-sampling and jump connection network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110674930A true CN110674930A (en) | 2020-01-10 |
Family
ID=69079699
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910927416.0A Pending CN110674930A (en) | 2019-09-27 | 2019-09-27 | SAR image denoising method based on learning down-sampling and jump connection network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110674930A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114173137A (en) * | 2020-09-10 | 2022-03-11 | 北京金山云网络技术有限公司 | Video coding method and device and electronic equipment |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110120020A (en) * | 2019-04-30 | 2019-08-13 | 西北工业大学 | A kind of SAR image denoising method based on multiple dimensioned empty residual error attention network |
| CN110147807A (en) * | 2019-01-04 | 2019-08-20 | 上海海事大学 | A kind of ship intelligent recognition tracking |
-
2019
- 2019-09-27 CN CN201910927416.0A patent/CN110674930A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110147807A (en) * | 2019-01-04 | 2019-08-20 | 上海海事大学 | A kind of ship intelligent recognition tracking |
| CN110120020A (en) * | 2019-04-30 | 2019-08-13 | 西北工业大学 | A kind of SAR image denoising method based on multiple dimensioned empty residual error attention network |
Non-Patent Citations (4)
| Title |
|---|
| KAI ZHANG等: "FFDNet：Toward a Fast and Flexible Solution for CNN Based Image Denoising", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
| 徐曾春等: "基于改进型WGAN的低剂量CT图像去噪方法", 《光学与光电技术》 * |
| 杜康宁等: "基于多层神经网络的中分辨SAR图像时间序列建筑区域提取", 《雷达学报》 * |
| 马德娇等: "基于卷积神经网络的SAR图像去噪方法", 《第五届高分辨率对地观测学术年会》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114173137A (en) * | 2020-09-10 | 2022-03-11 | 北京金山云网络技术有限公司 | Video coding method and device and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Singh et al. | Single image dehazing using gradient channel prior | |
| JP6057881B2 (en) | Method for removing noise from input image consisting of pixels containing noise | |
| Yang et al. | Y-net: Multi-scale feature aggregation network with wavelet structure similarity loss function for single image dehazing | |
| CN109035189B (en) | Infrared and weak visible light image fusion method based on Cauchy fuzzy function | |
| Bhandari et al. | Optimal sub-band adaptive thresholding based edge preserved satellite image denoising using adaptive differential evolution algorithm | |
| CN110517195B (en) | Unsupervised SAR image denoising method | |
| Sulam et al. | Image denoising through multi-scale learnt dictionaries | |
| CN103295204B (en) | A kind of image self-adapting enhancement method based on non-down sampling contourlet transform | |
| CN103020922A (en) | PCA (principal component analysis) transformation based SAR (synthetic aperture radar) image speckle suppression method | |
| Teng et al. | Modified pyramid dual tree direction filter‐based image denoising via curvature scale and nonlocal mean multigrade remnant filter | |
| CN104657951A (en) | Multiplicative noise removal method for image | |
| Kaur et al. | Image de-noising using wavelet transform and various filters | |
| Yao et al. | The Retinex-based image dehazing using a particle swarm optimization method | |
| CN103871031A (en) | Kernel regression-based SAR image coherent speckle restraining method | |
| CN110674930A (en) | SAR image denoising method based on learning down-sampling and jump connection network | |
| Santosh et al. | Efficient techniques for denoising of speckle and highly corrupted impulse noise images | |
| Tian et al. | Underwater sonar image denoising through nonconvex total variation regularization and generalized Kullback–Leibler fidelity | |
| Santoso et al. | Performance of various speckle reduction filters on Synthetic Aperture Radar image | |
| Wang et al. | An edge-preserving adaptive image denoising | |
| CN107945142B (en) | Synthetic aperture radar image denoising method | |
| Stolojescu-Crisan et al. | Denoising and inpainting SONAR images | |
| CN114240990A (en) | SAR image point target segmentation method | |
| Lu et al. | Application of improved CNN in SAR image noise reduction | |
| Bhargava et al. | An Effective Method for Image Denoising Using Non-local Means and Statistics based Guided Filter in Nonsubsampled Contourlet Domain. | |
| Xia et al. | Low SNR sonar image restoration based on mixed probability statistical model in wavelet domain |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20200110 |
|
| WD01 | Invention patent application deemed withdrawn after publication |