SAR image speckle noise processing method based on deep learning

Abstract

The invention discloses an SAR image speckle noise processing method based on deep learning, and the method comprises the steps: inputting an SAR image of real speckle noise into a speckle noise coding network, and carrying out the extraction of a speckle noise representation vector; and then, inputting the speckle noise representation vector and the input noiseless optical remote sensing image into a speckle noise reconstruction network, and transmitting the speckle noise of the input real SAR image to the input noiseless optical remote sensing image, thereby obtaining a real speckle noise degraded optical remote sensing image. And completing data set construction of the SAR image without the speckle noise and the SAR image with the speckle noise. The problem that a real SAR image without speckle noise is difficult to obtain in an SAR image speckle noise suppression method based on deep learning is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to a method for processing speckle noise in SAR images based on deep learning. Background Art

[0002] SAR images are a type of remote sensing images generated by a Synthetic Aperture Radar (SAR) system. As an active detection technology for earth observation, the synthetic aperture radar can be unaffected by darkness and bad weather conditions, and has the characteristics of all-weather and all-time imaging, and has been successfully applied in many fields such as military, mapping, geological detection, climate warning, forest fire prevention, disaster relief, etc.

[0003] However, due to the limitations of the synthetic aperture radar imaging principle, SAR images are severely affected by speckle noise, which has caused great obstacles to image interpretation and subsequent application work. Therefore, a large number of researchers have carried out corresponding work on the suppression of SAR image speckle. Currently, most of the mainstream methods use deep learning-based methods for research, and deep learning-based methods mostly use supervised learning techniques for training, and have a significant effect on the suppression of SAR image speckle noise. However, the supervised learning method usually needs to construct a training data pair of a SAR image without speckle damage and a SAR image damaged by speckle for training, and a SAR image without speckle damage cannot be obtained in principle because speckle noise will inevitably be introduced during the SAR image imaging process. Currently, the mainstream technical means mostly use synthetic data to construct the training data pair, convert the denoised clear visible light image into a grayscale image as the SAR image without speckle damage, and add synthetic speckle noise to the grayscale image as the SAR image damaged by speckle. However, the dataset constructed in this way cannot effectively simulate the distribution state of real speckle noise and the corresponding real speckle noise generation process, resulting in the inability to effectively process SAR images damaged by actual speckle to effectively suppress real speckle noise. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a more reasonable method for processing speckle noise. The speckle noise in the SAR image with real speckle noise is extracted by using a speckle noise encoding network, and then the extracted speckle noise is transmitted to an optical remote sensing image by using a speckle noise reconstruction network to generate a SAR image damaged by speckle, so that the synthesized SAR image damaged by speckle is more in line with the characteristics of real speckle noise.

[0005] Using the above speckle noise processing method, a highly realistic image is obtained, and the processed image is selected according to different geological characteristics and then a training set is established, and the corresponding supervised learning model achieves a good speckle noise suppression result.

[0006] To achieve the above technical objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for processing SAR image speckle noise based on deep learning includes the following steps:

[0008] 1) Convert the SAR image with real speckle noise into a high-dimensional deep feature vector.

[0009] 2) Pass the high-dimensional deep feature vector through multiple convolutional layers and activation functions and then input it into the average pooling layer for feature reduction, and then pass it through multiple convolutional layers and activation functions again to obtain a speckle noise representation vector.

[0010] 3) Convert the optical remote sensing image into a high-dimensional shallow feature vector.

[0011] 4) Adjust the dimension of the speckle noise representation vector and calculate the modulated weight.

[0012] 5) Input the modulated weight and the high-dimensional shallow feature vector into the demodulation module to obtain a demodulated feature vector.

[0013] 6) Input the demodulated feature vector into a convolutional layer and an activation function to obtain an optical remote sensing feature vector containing speckle noise.

[0014] 7) Replace the high-dimensional shallow feature vector with the optical remote sensing feature vector containing speckle noise and repeat steps 4)-6) for a predetermined number of times, and then output an optical remote sensing image feature vector with sufficient speckle noise transmission.

[0015] 8) Input the optical remote sensing image feature vector with sufficient speckle noise transmission into a convolutional layer to obtain an image with real speckle noise.

[0016] As one of the preferred solutions, in step 1), a network with a UNet structure is used to obtain a high-dimensional deep feature vector.

[0017] As one of the preferred solutions, in step 2), the high-dimensional deep feature vector is passed through 3 convolutional layers and activation functions and then input into the average pooling layer for feature reduction, and then passed through 3 convolutional layers and activation functions again to obtain a speckle noise representation vector.

[0018] As one of the preferred solutions, step 4) further includes a step of reducing and optimizing the shallow feature vector.

[0019] As one of the preferred solutions, the method for refining and optimizing the shallow feature vector FV1 in step 4) is as follows:

[0020] FB1 = relu(conv8(FV1))

[0021] FB2 = FV1(conv9(FB1))

[0022] Among them, conv8 and conv9 are respectively 3*3 convolution operations, relu is an activation function in torch.nn, and FB1 and FB2 are respectively the optical remote sensing feature vectors after successive refinement and optimization.

[0023] As one of the preferred solutions, the method for modulating the speckle noise representation vector E after dimension adjustment in step 4) is as follows:

[0024]

[0025] Among them, lw is a learnable weight parameter, is a scale quantization parameter, and ω is the weight after the modulation module modulates the speckle noise representation vector E.

[0026] As one of the preferred solutions, the modulated weight ω and the refined and optimized feature vector FB2 in step 4) are input into the following demodulation module:

[0027]

[0028] Among them, psum 2,3,4 is an operation of summing the 2nd, 3rd, and 4th dimensions of the vector respectively, and FB3 is the demodulated feature vector output by the demodulation module after noise transfer.

[0029] As one of the preferred solutions, step 8) also includes an optimization processing step of performing a convolution layer and an activation function on the optical remote sensing image feature vector with sufficient speckle noise transfer.

[0030] The advantages and beneficial effects of the present invention are as follows:

[0031] The present invention uses a speckle noise encoding network to extract a noise representation vector from a SAR image with real speckle noise, and then uses this noise representation vector in a speckle noise reconstruction network to process the input optical remote sensing image to obtain a synthetic image damaged by real speckle noise, solving the problem that it is difficult to obtain a real SAR image without speckle noise in the method for suppressing speckle noise in SAR images based on deep learning. First, the SAR image with real speckle noise is input into the speckle noise encoding network to extract the speckle noise representation vector, which represents the severity, spatial distribution state, and corresponding potential degradation model of the speckle noise of the input real SAR image. Then, the speckle noise representation vector and the input noise-free optical remote sensing image are input into the speckle noise reconstruction network, and the speckle noise of the input real SAR image is transferred to the input noise-free optical remote sensing image, so as to obtain an optical remote sensing image degraded by real speckle noise, and complete the construction of a data set of SAR images without speckle noise and SAR images with speckle noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic flow chart of the present invention.

[0033] Figure 2 It is a schematic diagram of the speckle noise encoding network.

[0034] Figure 3 It is a schematic diagram of the speckle noise reconstruction network.

[0035] Figure 4 It is a schematic diagram of the vector guidance generation module.

[0036] Figure 5 It is a diagram of the result of real speckle noise transfer.

[0037] For those of ordinary skill in the art, other related drawings can be obtained based on the above drawings without creative efforts. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] In order to enable those skilled in the art of the present technology to better understand the solution of the present invention, the technical solution of the present invention will be further described below in conjunction with the drawings and specific embodiments.

[0039] The method for processing speckle noise in SAR images based on deep learning of the present invention includes the following steps:

[0040] 1) Convert the SAR image with real speckle noise into a high-dimensional deep feature vector.

[0041] 2) Input the high-dimensional deep feature vectors into an average pooling layer for feature reduction after passing through multiple convolutional layers and activation functions, and then obtain the speckle noise representation vectors again through multiple convolutional layers and activation functions.

[0042] 3) Convert the optical remote sensing image into high-dimensional shallow feature vectors, preferably using noise-free optical remote sensing images.

[0043] 4) Adjust the dimension of the speckle noise representation vectors and calculate the modulated weights.

[0044] 5) Input the modulated weights and high-dimensional shallow feature vectors into a demodulation module to obtain the demodulated feature vectors.

[0045] 6) Input the demodulated feature vectors into convolutional layers and activation functions to obtain optical remote sensing feature vectors containing speckle noise.

[0046] 7) Replace the high-dimensional shallow feature vectors with the optical remote sensing feature vectors containing speckle noise, and then repeat steps 4)-6) for a predetermined number of times and output the optical remote sensing image feature vectors with sufficient transmission of speckle noise.

[0047] 8) Input the optical remote sensing image feature vectors with sufficient transmission of speckle noise into a convolutional layer to obtain an image with real speckle noise.

[0048] In the prior art, it is simply considered that speckle noise conforms to the gamma distribution. Numerically generate values of the gamma distribution with the same size as the image, and then multiply them with the optical remote sensing grayscale image to obtain a synthetic SAR image damaged by speckle noise. The main difference between the synthesis here and the synthesis of this method is that the synthetic speckle noise in the traditional method is only multiplicative noise modeled by the gamma distribution and cannot truly reflect the real degraded speckle noise; the present invention uses a deep learning method to transfer the speckle noise on real SAR images to optical remote sensing images, which can be more in line with the real speckle noise distribution.

[0049] The present invention proposes a more reasonable method for constructing a synthetic dataset of speckle noise. Use a speckle noise encoding network to extract the speckle noise from the SAR image with real speckle noise, and then use a speckle noise reconstruction network to transfer the extracted speckle noise to the optical remote sensing image to generate a SAR image damaged by speckle, making the synthetic SAR image damaged by speckle more in line with the characteristics of real speckle noise and enabling the corresponding supervised learning model to achieve better speckle noise suppression results.

[0050] The entire process will be described below in two parts: speckle noise extraction and reconstruction. The speckle noise encoding network is used to extract speckle noise, and the speckle reconstruction network is used to reconstruct speckle noise. The following is a specific description.

[0051] 1. Speckle Noise Encoding Network

[0052] First, use the speckle noise encoding network to extract the speckle noise representation vector from the SAR image with real speckle noise, as shown below:

[0053] E = Encoder(I N ) (1)

[0054] where E is the extracted speckle noise representation vector, which is a one-dimensional vector with a size of 512×1×1. I N is the SAR image with real speckle noise, Encoder is the speckle noise encoding network, and the detailed structure is shown in Figure 2 .

[0055] As one implementation, the input SAR image I with real speckle noise is converted into a high-dimensional deep feature vector through the deep feature extraction module. The deep feature extraction module is a network with a classic UNet structure. The corresponding structure information is as follows: the network structure has 4 layers, the number of corresponding encoding blocks is 2, 2, 4, 8 respectively, the number of corresponding decoding blocks is 8, 4, 2, 2 respectively, and the number of modules in the middle layer is 16. Here, only the above introduction of this classic UNet network used in the present invention is given, and no further details will be elaborated. Through the deep feature extraction module of the UNet structure, I with a size of 1×H×W is N converted into a deep feature vector with a size of nf×H×W, as shown below: N

[0056] F1 = UNet(I N ) (2)

[0057] where UNet is the deep feature extraction module, F1 is the extracted deep feature vector, H and W are the height and width of the input image F1 respectively. The 1 in 1×H×W is the number of channels of the input SAR image, which is 1, and nf in nf×H×W is the vector dimension of the deep feature vector F1. In the present invention, nf is set to 64. The larger the dimension, the larger the model and the higher the computational complexity. It can also be set to 32, but this will reduce the model performance. Setting it to 64 is a trade-off between performance and computational cost.

[0058] Then, the extracted deep feature vector F1 is input into the first convolutional layer and the activation function to obtain the second deep feature vector F2, as shown below: ​

[0059] F2 = LeakyReLU(conv1(F1)) (3)

[0060] Where LeakyReLU is an activation function in torch.nn. Of course, other suitable activation functions can also be selected. The same applies hereinafter. conv1 is a 3×3 convolution operation. After passing through the first convolutional layer and the activation function, the size of the second deep feature vector F2 becomes 128×(H / 2)×(W / 2).

[0061] Continue to operate on the second deep feature vector F2, input it into the second convolutional layer and the activation function, and obtain the third deep feature vector F3 as follows:

[0062] F3 = LeakyReLU(conv2(F2)) (4)

[0063] Here, LeakyReLU is an activation function in torch.nn, and conv2 is a 3×3 convolution operation. After passing through the second convolutional layer and the activation function, the size of the third deep feature vector F3 becomes 256×(H / 4)×(W / 4).

[0064] Continue to operate on the third deep feature vector F3, input it into the third convolutional layer and the activation function, and obtain the fourth deep feature vector F4 as follows:

[0065] F4 = LeakyReLU(conv3(F3)) (5)

[0066] Here, LeakyReLU is an activation function in torch.nn, and conv2 is a 3×3 convolution operation. After passing through the third convolutional layer and the activation function, the size of the fourth deep feature vector c becomes 512×(H / 8)×(W / 8). After passing through the above series of convolution operations, input the fourth deep feature vector F4 into the average pooling layer for feature reduction, as follows:

[0067] F5 = AdaptiveAvgPool2d(F4) (6)

[0068] Where F5 is the fifth layer feature vector after feature reduction, and AdaptiveAvgPool2d is a pooling function in torch.nn. After passing through the average pooling layer, the size of the fifth layer feature vector F5 becomes 512×1×1. Then, input the fifth layer feature vector F5 after feature reduction into the fourth convolutional layer and the activation function, and obtain the sixth deep feature vector F6 as follows:

[0069] F6 = LeakyReLU(conv4(F5)) (7)

[0070] Among them, LeakyReLU is an activation function in torch.nn, and conv4 is a 1×1 convolution operation. After passing through convolutional layer four and the activation function, the size of the sixth deep feature vector F6 remains unchanged. Continuing to operate on the sixth deep feature vector F6, it is input into convolutional layer five and the activation function to obtain the seventh deep feature vector F7, as shown below:

[0071] F7 = LeakyReLU(conv5(F6)) (8)

[0072] Among them, LeakyReLU is an activation function in torch.nn, and conv5 is a 1×1 convolution operation. After passing through convolutional layer five and the activation function, the size of the seventh deep feature vector F7 remains unchanged. Finally, the seventh deep feature vector F7 is input into convolutional layer six to obtain the extracted speckle noise representation vector E, specifically as follows:

[0073] E = conv6(F7) (9)

[0074] Among them, conv6 is a 1×1 convolution operation, E is the extracted speckle noise representation vector, and the vector size is 512×1×1.

[0075] 2. Speckle Noise Reconstruction Network

[0076] Through the above steps, the true speckle noise representation vector is extracted, which can fully express the distribution state, degree, and the corresponding true speckle noise generation process of the noise. Then, the speckle noise reconstruction network is used to complete the transfer of the true speckle noise using the speckle noise representation vector, transferring the true speckle noise from the original SAR image to the clean and noiseless remote sensing image to complete the construction of a more reasonable synthetic dataset. The specific process is as Figure 3 shown.

[0077] First, the SAR image is a single-channel grayscale image. For optical remote sensing images, they need to be converted to grayscale images. Visible light images are RGB three-channel images, which are converted to single-channel grayscale images to simulate SAR images. Then, the optical remote sensing image is input into the shallow feature extraction module to convert the original optical remote sensing image into a high-dimensional shallow feature vector, specifically as follows:

[0078] FV1 = conv7(I V ) (10)

[0079] Among them, conv7 is a 3×3 convolution operation, FV1 is the extracted high-dimensional shallow feature vector, and I V is the input original noiseless optical remote sensing image.

[0080] Next, the extracted high-dimensional shallow feature vector FV1 and the extracted speckle noise representation vector E are sequentially input into four representation vector-guided generation modules to fully transfer the speckle noise to the feature vector of the optical remote sensing image, as follows:

[0081] FM1 = GVGB1(FV1, E) (11)

[0082] FM2 = GVGB2(FM1, E) (12)

[0083] FM3 = GVGB3(FM2, E) (13)

[0084] FM4 = GVGB4(FM3, E) (14)

[0085] Among them, GVGB1, GVGB2, GVGB3, and GVGB4 are respectively the same representation vector-guided generation modules, and FM1, FM2, FM3, and FM4 are respectively the optical remote sensing feature vectors containing speckle noise sequentially generated by the corresponding GVGB1, GVGB2, GVGB3, and GVGB4 modules. Here, the structures of the four representation vector-guided generation modules described above are the same. Only one module GVGB1 is described in detail, and the corresponding structure diagram is as Figure 4 shown.

[0086] After the high-dimensional shallow feature vector FV1 and the extracted speckle noise representation vector E are input into the representation vector-guided generation module GVGB1, the shallow feature vector FV1 is first input into the feature reduction module to obtain the reduced and optimized feature vector, as follows:

[0087] FB1 = relu(conv8(FV1)) (15)

[0088] FB2 = FV1(conv9(FB1)) (16)

[0089] Among them, conv8 and conv9 are respectively 3*3 convolution operations, relu is the activation function in torch.nn, and FB1 and FB2 are respectively the optical remote sensing feature vectors after successive reduction and optimization.

[0090] Then, the speckle noise representation vector E is input into the modulation module for noise vector modulation operation. First, a dimensional expansion operation is performed on the speckle noise representation vector E, expanding it from 512×1×1 to 1×512×1×1. Then, the expanded speckle noise representation vector E is operated on to complete the modulation operation, as follows:

[0091]

[0092] Among them, lw is a learnable weight parameter, and its initial value is set to a random number between 0 and 1 with a size of 64×64×3×3. For example, it can be initially set to 1, and the value of the parameter can be updated during backpropagation in the training process and gradually become optimal as the training process progresses. It is a scale quantization parameter, and its value is determined by the input feature vector and the size of the selected convolution kernel. The convolution kernel size used in the present invention is 3×3, and the vector dimension is 64. Therefore, 3×3×64 = 576.

[0093] ω is the weight after the modulation module modulates the speckle noise representation vector E. It is set according to the size of the selected convolution kernel and the dimension of the input feature vector. Then, the modulated weight ω and the refined and optimized feature vector FB2 are input into the demodulation module for the transfer operation of speckle noise. Specifically:

[0094]

[0095] Among them, psum 2,3,4 is the summation operation for the second, third, and fourth dimensions of the vector respectively, and FB3 is the feature vector after noise transfer output by the demodulation module.

[0096] After the demodulated feature vector FB3 is sequentially input into the convolutional layer and the activation function, the final representation vector is obtained to guide the optical remote sensing feature vector FM1 containing speckle noise output by the generation module GVGB1, as follows:

[0097] FM1 = LeakyReLU(conv10(FB3)) (19)

[0098] Among them, conv10 is a 3×3 convolutional operation, and LeakyReLU is an activation function in torch.nn.

[0099] The above is the entire process of the representation vector guiding generation module GVGB1. Then, FM1 is processed by the same representation vector guiding generation module three times according to formulas (11)-(14) to obtain the optical remote sensing image feature vector FM4 with sufficient speckle noise transfer. In order to fully transfer the noise, the purpose of setting 4 times is a trade-off between the amount of calculation and the model performance. Setting too few will result in a decrease in performance, and setting too many will increase the amount of calculation.

[0100] Next, a convolutional layer and an activation function are successively executed on FM4, as follows:

[0101] FM5 = LeakyReLU(conv11(FM4)) (20)

[0102] Among them, conv11 is a 3×3 convolution operation, LeakyReLU is an activation function in torch.nn, and FM5 is the feature vector of an optical remote sensing image with real speckle noise after optimization processing. Finally, FM5 is input into a convolutional layer to obtain the final optical remote sensing image with real speckle noise, as follows:

[0103] S = conv12(FM5) (21)

[0104] Among them, conv12 is a 3×3 convolution operation, and S is the final optical remote sensing image with real speckle noise.

[0105] As Figure 5 shown, the three images in the left column show the optical remote sensing images without noise, the three images in the middle column are SAR images with real coherent noise, and the three images in the right column are the optical remote sensing images after speckle noise transfer processed by the method of the present invention. Among them, the attached drawings only show a few pieces of data. The real generated data will use a large number of real SAR images to generate different types of data, which will include various terrain characteristics. The terrain characteristics include natural characteristics, such as mountains, deserts, towns, and grasslands, etc., and will also include some dynamically changing terrain geology, such as sea surface pollution, plant dynamic growth stages, etc. Supplementary with a large number of initial pictures with different characteristics, after being processed by the method of the present invention, it will have better applicability, and using this as a training set can improve the training effect of the model.

[0106] To further improve the processing effect, for different environments, environmental features are added as additional inputs in each encoding and decoding block, skip connections are added, and the environmental features are directly passed to the corresponding decoding block. At the same time, an adaptive feature fusion layer can be set to dynamically adjust the weights of the UNet output features and environmental features, and the attention mechanism is used to achieve adaptive fusion of features.

[0107] For example, in the polar ice and snow monitoring scenario, SAR images are often used to monitor glacier changes and sea ice coverage. However, SAR images in the polar environment are often affected by more complex speckle noise, which is caused by the special atmospheric conditions, temperature changes, and surface reflection characteristics in the polar region. In this case, traditional speckle noise coding networks may not be able to effectively capture the special noise patterns in polar SAR images.

[0108] Furthermore, an adaptive special terrain, such as a polar speckle noise coding network, is proposed. Based on the original UNet structure, this network adds a special environment feature extraction module, such as a polar environment feature extraction module. This module contains multiple parallel spatial attention mechanisms, each of which is specifically used to extract different types of polar environment features (such as ice surface texture, snow layer depth, etc.). The speckle noise representation vector processed by the spatial attention mechanism then has the specificity corresponding to this terrain. The training set established for this terrain feature after processing such images can reflect higher authenticity. In specific implementation, a simple spatial attention module can usually be constructed using convolutional layers, pooling layers, and activation functions. Then, multiple spatial attention modules are applied in parallel to the input initial SAR image, and then the output feature maps are fused to obtain multiple feature maps enhanced by attention. The fusion methods can adopt different fusion methods such as splicing fusion, addition fusion, weighted addition fusion, etc. Different fusion methods can be selected according to the situation. For example, in the case where a certain adaptive ability is required and the importance of different attention mechanisms needs to be learned, and the importance of different attention mechanisms is uncertain, the weighted addition fusion method can be selected.

[0109] The above makes an exemplary description of the present invention. It should be noted that without departing from the core of the present invention, any simple deformation, modification, or equivalent replacement that can be made by those skilled in the art without creative labor falls within the protection scope of the present invention.

Claims

1. A method for processing speckle noise in SAR images based on deep learning, characterized in that, Including the following steps, 1) Convert the SAR image with real speckle noise into a high-dimensional deep feature vector, 2) Input the high-dimensional deep feature vector into the average pooling layer for feature reduction after passing through multiple convolutional layers and activation functions, and then obtain the speckle noise representation vector by passing through multiple convolutional layers and activation functions again, 3) Convert the optical remote sensing image into a high-dimensional shallow feature vector, 4) Adjust the dimension of the speckle noise representation vector and calculate the modulated weight, 5) Input the modulated weight and the high-dimensional shallow feature vector into the demodulation module to obtain the demodulated feature vector, 6) Input the demodulated feature vector into the convolutional layer and activation function to obtain the optical remote sensing feature vector containing speckle noise, 7) Replace the high-dimensional shallow feature vector with the optical remote sensing feature vector containing speckle noise, and then repeat steps 4)-6) for a predetermined number of times to output the optical remote sensing image feature vector with sufficient speckle noise transmission, 8) Input the optical remote sensing image feature vector with sufficient speckle noise transmission into a convolutional layer to obtain an image with real speckle noise.

2. The method for processing speckle noise in SAR images based on deep learning according to claim 1, wherein In step 1), a network with a UNet structure is used to obtain the high-dimensional deep feature vector.

3. The method for processing speckle noise in SAR images based on deep learning according to claim 1, characterized in that, In step 2), the high-dimensional deep feature vector is input into the average pooling layer for feature reduction after passing through 3 convolutional layers and activation functions, and then the speckle noise representation vector is obtained by passing through 3 convolutional layers and activation functions again.

4. The method for processing speckle noise in SAR images based on deep learning according to claim 1, characterized in that, Step 4) also includes a step of reducing and optimizing the shallow feature vector.

5. The method for processing speckle noise in SAR images based on deep learning according to claim 4, characterized in that, The method for reducing and optimizing the shallow feature vector FV1 in step 4) is, FB1 = relu(conv8(FV1)) FB2 = FV1(conv9(FB1)) where conv8 and conv9 are respectively 3×3 convolutional operations, relu is the activation function in torch.nn, and FB1 and FB2 are respectively the optical remote sensing feature vectors after successive reduction and optimization.

6. The method for processing speckle noise in SAR images based on deep learning according to claim 5, characterized in that, The method for modulating the dimension-adjusted speckle noise representation vector E in step 4) is, where, lw is a learnable weight parameter, is a scale quantization parameter, and ω is the weight after the modulation module modulates the speckle noise representation vector E.

7. The method for processing speckle noise in SAR images based on deep learning according to claim 6, wherein In step 4), the modulated weight ω and the reduced and optimized feature vector FB2 are input into the following demodulation module, where psum 2,3,4 is the summation operation on the 2nd, 3rd, and 4th dimensions of the vector, and FB3 is the eigenvector after demodulation and noise transfer output by the demodulation module.

8. The method for processing speckle noise in SAR images based on deep learning according to claim 1, characterized in that, Step 8) also includes an optimization processing step of performing a convolutional layer and activation function on the optical remote sensing image feature vector with sufficient speckle noise transmission.

Images