SAR image speckle noise suppression method based on multi-scale strong complex domain alternating restoration network

Abstract

The application relates to the field of image noise suppression and discloses a SAR image speckle noise suppression method based on a multi-scale strong complex domain alternating restoration network, which comprises the following steps: acquiring an input SAR image with speckle noise and extracting a shallow high-dimensional feature vector thereof; constructing and pre-training a multi-scale strong complex domain alternating restoration network; the multi-scale strong complex domain alternating restoration network comprises an encoding part and a decoding part; inputting the shallow high-dimensional feature vector into the encoding part for speckle noise suppression to obtain a processed image; inputting the processed image into the decoding part to obtain high-frequency detail restoration and obtain a final SAR image with speckle noise suppression; and the application can realize high-fidelity restoration of the SAR image.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image noise suppression, and particularly relates to a SAR image speckle noise suppression method based on a multi-scale strong complex domain alternating restoration network. BACKGROUND

[0002] Synthetic aperture radar is a common technical means for ground observation, and the image information (SAR image) obtained thereby has a wide range of applications in various industries. Moreover, compared with optical remote sensing images, SAR images have the advantages of all-weather and all-day due to their active emission characteristics, and thus have irreplaceable characteristics and play an important role in black night and adverse weather conditions.

[0003] The original data obtained by synthetic aperture radar for ground observation is generally in complex format, and the real part and the imaginary part are squared and summed to obtain the intensity format SAR image after multi-view processing. However, there is usually severe fluctuation of speckle noise in the intensity format SAR image, which seriously affects the development of subsequent application work. Researchers usually suppress the speckle noise based on the intensity format SAR image, and they consider the speckle noise as a kind of compound gamma distribution multiplicative noise for processing. Based on the above noise modeling process, a large number of research methods are proposed, and good speckle noise suppression results are obtained. However, it is difficult to effectively restore the real and undamaged SAR image by suppressing the speckle noise in the intensity domain only, because the intensity format SAR image is obtained by squaring and summing the real part and the imaginary part respectively, and it is difficult to accurately restore the real and imaginary part information in the intensity domain only, so that the SAR image cannot be restored with high fidelity. SUMMARY

[0004] The purpose of the present application is to provide a SAR image speckle noise suppression method based on a multi-scale strong complex domain alternating restoration network, which solves the technical problem that the existing method cannot restore the SAR image with high fidelity.

[0005] Specifically, the present application provides a SAR image speckle noise suppression method based on a multi-scale strong complex domain alternating restoration network, which uses an alternatingly connected intensity domain speckle noise suppression module and a complex domain real part and imaginary part restoration module to suppress the speckle noise and then convert it to the complex domain to further accurately restore the real part and imaginary part information. The network main body of the UNet structure is used in the encoding part to perform the above speckle suppression, real part and imaginary part information restoration work at multiple scales to fully suppress the speckle noise, and then the dilated convolution attention module and the high-frequency detail restoration module are used in the decoding part of the UNet structure to increase additional attention to the high-frequency texture part to further restore it to obtain a high-fidelity restoration result.

[0006] Specifically, the method comprises the following steps:

[0007] S1, acquire an input SAR image with coherent speckle noise, and extract a shallow high-dimensional feature vector thereof;

[0008] S2, construct and pre-train a multi-scale strong complex domain alternating restoration network; the multi-scale strong complex domain alternating restoration network comprises an encoding part and a decoding part;

[0009] S3, input the shallow high-dimensional feature vector to the encoding part for coherent speckle noise suppression, to obtain a processed image;

[0010] S4, input the processed image to the decoding part to obtain high-frequency detail restoration, to obtain a final SAR image with coherent speckle noise suppressed.

[0011] A storage device stores instructions and data for implementing a SAR image coherent speckle noise suppression method based on a multi-scale strong complex domain alternating restoration network.

[0012] A SAR image coherent speckle noise suppression device based on a multi-scale strong complex domain alternating restoration network comprises a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a SAR image coherent speckle noise suppression method based on a multi-scale strong complex domain alternating restoration network.

[0013] The present application provides the beneficial effect that a SAR image coherent speckle noise suppression method based on a multi-scale strong complex domain alternating restoration network is proposed; the network body adopts a three-layer UNet structure, the encoding part thereof focuses on coherent speckle noise suppression, and the decoding part thereof focuses on high-frequency detail restoration and boundary feature sharpening; the encoding part uses alternatingly connected strength domain coherent speckle noise suppression modules and complex domain real part and imaginary part restoration modules to alternately perform coherent speckle noise suppression and complex domain restoration operations to achieve coherent speckle noise suppression without losing the amplitude and phase information contained; the decoding part uses a cavity convolution attention module and a high-frequency detail restoration module to add additional attention to the high-frequency texture part to restore it, to obtain a SAR image with coherent speckle noise suppressed, and finally enable the present application to restore the SAR image with high fidelity. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is a multi-scale strong complex domain alternating restoration network processing flowchart of the present application;

[0015] Figure 2 is a strength domain coherent speckle noise suppression module structure schematic diagram;

[0016] Figure 3 is a complex domain real part and imaginary part restoration module structure schematic diagram;

[0017] Figure 4 This is a schematic diagram of the dilated convolutional attention module structure;

[0018] Figure 5 This is a schematic diagram of the high-frequency detail restoration module structure;

[0019] Figure 6 This is a schematic diagram of speckle noise suppression results based on a multi-scale strong complex domain alternating restoration network;

[0020] Figure 7 This is a schematic diagram of the hardware device used in this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0022] Before formally describing the present invention, a general description of the solution of the present invention will be given first to facilitate understanding.

[0023] Please refer to Figure 1 This invention provides a method for suppressing speckle noise in SAR images based on a multi-scale strong complex domain alternating restoration network, comprising the following steps:

[0024] S1. Obtain the input SAR image with speckle noise and extract its shallow high-dimensional feature vector;

[0025] It should be noted that, firstly, shallow high-dimensional feature vectors are extracted from the input SAR image with speckle noise to fully express the relevant information of the SAR image, as shown below:

[0026] (1)

[0027] in, , The input SAR image with speckle noise and the shallow high-dimensional feature vector extracted by the convolutional layer are respectively. conv1 is a 3×3 convolution operation with 1 input channel and 64 output channels.

[0028] S2. Construct and pre-train a multi-scale strong complex domain alternating reconstruction network; the multi-scale strong complex domain alternating reconstruction network includes an encoding part and a decoding part;

[0029] It should be noted that the encoding part in step S2 includes: an intensity domain speckle noise suppression module, a complex domain real and imaginary part restoration module, and a downsampling module.

[0030] Specifically, after extracting the shallow high-dimensional feature vectors, they are input into a three-layer UNet network to suppress speckle noise, restore high-frequency details, and sharpen boundary features. The overall method flowchart is shown below. Figure 1 As shown.

[0031] The main body of the network can be divided into an encoding part and a decoding part. The encoding part focuses on suppressing speckle noise, while the decoding part focuses on restoring high-frequency details and sharpening boundary features.

[0032] S3. Input the shallow high-dimensional feature vector into the encoding part to suppress speckle noise and obtain the processed image;

[0033] It should be noted that step S3 is as follows:

[0034] S31. Input the shallow high-dimensional feature vector into the intensity domain speckle noise suppression module to initially suppress speckle noise in the intensity domain at the first scale;

[0035] S32. The feature vector after suppressing speckle noise In the module for restoring the real and imaginary parts of the input complex number, information is restored for the real and imaginary parts in the complex number domain before calculating the intensity domain feature vector. ;

[0036] S33. The restored intensity domain feature vector Feature vector after suppressing speckle noise The input is fed into the downsampling module to further suppress speckle noise at the second scale and restore the real and imaginary parts of the information in the complex domain, resulting in the feature vector after speckle noise suppression at the second scale. eigenvectors of the restored intensity domain ;

[0037] S34. Perform a 2x downsampling and repeat step S33 above to obtain a high-dimensional vector after fully suppressing speckle noise at the third scale. Obtain the restored intensity domain feature vector of the encoded part output, including the feature vector after speckle noise suppression at the third scale. eigenvectors of the restored intensity domain RF 3.

[0038] As one example, the encoding part will be described in detail below. The encoding part mainly uses three types of modules: intensity domain speckle noise suppression module, complex domain real and imaginary part restoration module, and downsampling module.

[0039] First, the extracted shallow high-dimensional feature vector is input into the intensity domain speckle noise suppression module to initially suppress speckle noise in the intensity domain at the first scale, as shown below:

[0040] (2)

[0041] in, The first-scale intensity-domain speckle noise suppression module represents the feature vector after suppressing speckle noise. IDSB is the first-scale intensity-domain speckle noise suppression module, and its structural diagram is shown below. Figure 2 As shown, details will be provided in subsequent steps.

[0042] Then, the feature vector after suppressing speckle noise... In the module for restoring the real and imaginary parts of the input complex number, information is restored from the real and imaginary parts in the complex number domain, and then the intensity domain eigenvector is calculated as the output of this module, as shown below:

[0043] (3)

[0044] in, The intensity domain feature vector is the result of the complex domain real and imaginary part restoration module at the first scale. CFRB is the complex domain real and imaginary part restoration module at the first scale, and its structural diagram is shown below. Figure 3 As shown, details will be provided in subsequent steps.

[0045] Next, the restored intensity domain feature vectors are analyzed. Downsampling is then performed to further suppress speckle noise at the second scale and restore the real and imaginary parts of the information in the complex domain, as shown below:

[0046] (4)

[0047] (5)

[0048] in, , These are the feature vectors of the intensity domain speckle noise suppression module and the feature vector of the intensity domain after restoration, respectively, output by the real and imaginary part restoration module in the intensity domain at the second scale. This refers to the `torch.nn.PixelUnshuffle` function in PyTorch used to perform downsampling operations. This is a 2x downsampling operation.

[0049] Repeat the above steps by performing a 2x downsampling again to obtain a high-dimensional vector with speckle noise fully suppressed at the third scale. The feature vector output by the encoding module is then obtained, as shown below:

[0050] (6)

[0051] (7)

[0052] in, , These are the eigenvectors of the intensity domain speckle noise suppression module and the eigenvectors of the intensity domain after restoration, respectively, output by the intensity domain speckle noise suppression module and the complex domain real and imaginary part restoration module.

[0053] S4. Input the processed image into the decoding section to obtain high-frequency detail restoration and the final SAR image with speckle noise suppression.

[0054] It should be noted that the decoding part in step S2 includes: a dilated convolutional attention module and a high-frequency detail restoration module.

[0055] It should be noted that step S4 is as follows:

[0056] S41. Input the restored intensity domain feature vector output from the encoding part into the dilated convolutional attention module and the high-frequency detail restoration module to perform high-frequency detail restoration and boundary feature sharpening, and obtain the feature vector after attention processing at the third scale. DCF 3. Feature vectors after high-frequency detail restoration ;

[0057] S42. The feature vector after restoring high-frequency details at the third scale. After upsampling, feature information at the second scale is obtained, and then the feature vector of the restored intensity domain output from the second scale of the encoding part is added. The input is fed into the dilated convolutional attention module and the high-frequency detail restoration module to restore the feature information at the second scale of the decoding part, resulting in the feature vector after attention processing at the second scale. DCF 2. Feature vectors after high-frequency detail restoration ;

[0058] S43. Perform upsampling again by 2x, repeating step S42 to obtain the feature vector for high-frequency detail restoration and boundary feature sharpening at the first scale, including... DCF 1 and ;

[0059] S44. Feature vectors after restoring high-frequency details at the first scale. Adding the shallow high-dimensional feature vectors extracted at the beginning of the network The input is fed into a contiguous dilated convolutional attention module and a high-frequency detail restoration module for the final high-frequency texture restoration step, resulting in a further processed feature vector from the outputs of the dilated convolutional attention module and the high-frequency detail restoration module. DCF 0. Feature vectors after high-frequency detail restoration .

[0060] Step S4 also includes:

[0061] S45. The high-frequency detail restoration feature vector output by the last high-frequency detail restoration module. The input is fed into a convolutional layer, which restores it from the vector domain to the image domain, resulting in the final SAR image with speckle noise suppressed.

[0062] As one embodiment, the encoding part of the above steps is considered to have obtained a high-dimensional vector that has sufficiently suppressed speckle noise. Then, it is input into the decoding part, and the dilated convolutional attention module and the high-frequency detail restoration module are used to add extra attention to the high-frequency texture part to restore it, so as to obtain the SAR image after speckle noise suppression.

[0063] Specifically, the restored intensity domain feature vectors output by the complex domain real and imaginary part restoration module at the third scale are input into the dilated convolutional attention module and the high-frequency detail restoration module to perform high-frequency detail restoration and boundary feature sharpening, as shown below:

[0064] (8)

[0065] (9)

[0066] Among them, IDCB and CHRB are the dilated convolutional attention module and the high-frequency detail restoration module at the third scale in the decoding module, respectively, and their structural diagrams are shown below. Figure 4 , Figure 5 As shown, the relevant content of the module will be described in detail in subsequent steps; , These are the feature vectors after attention processing obtained from IDCB and CHRB processing at the third scale, and the feature vectors after high-frequency detail restoration, respectively.

[0067] Next, the feature vectors after high-frequency detail restoration at the third scale will be... After upsampling, feature information at the second scale is obtained, and then the feature vector of the restored intensity domain output from the second scale of the encoding part is added. The input is fed into the dilated convolutional attention module and the high-frequency detail restoration module to restore the feature information at the second scale of the decoding part, as shown below:

[0068] (10)

[0069] (11)

[0070] in, , These represent the feature vectors after attention processing at the second scale, output by the dilated convolutional attention module, the high-frequency detail restoration module, and the high-frequency detail restoration module, respectively. This refers to the `torch.nn.Pixelshuffle` function in PyTorch used to perform downsampling operations. This is a 2x upsampling operation.

[0071] Perform upsampling again by a factor of 2, and combine it with the feature vector of the restored intensity domain output from the first scale of the encoding part. Repeating the above steps yields the feature vectors for high-frequency detail restoration and boundary feature sharpening at the first scale, as shown below:

[0072] (12)

[0073] (13)

[0074] in, , These are the feature vectors after attention processing at the first scale output by the dilated convolutional attention module and the high-frequency detail restoration module, respectively, and the feature vectors after high-frequency detail restoration.

[0075] Then, the feature vector after restoring the high-frequency details at the first scale is... Adding the shallow high-dimensional feature vectors extracted at the beginning of the network The input is fed into a continuously connected dilated convolutional attention module and a high-frequency detail restoration module for the final high-frequency texture restoration step, as shown below:

[0076] (14)

[0077] (15)

[0078] in, , These are the feature vectors after attention processing and the feature vectors after high-frequency detail restoration, respectively, output by the dilated convolution attention module, the high-frequency detail restoration module, and the high-frequency detail restoration module.

[0079] Finally, the high-frequency detail restoration feature vector output by the last high-frequency detail restoration module is... The input is fed into a convolutional layer, which restores the image from the vector domain to the image domain, resulting in the SAR image with speckle noise suppression output by this network, as shown below:

[0080] (16)

[0081] in, This is the feature vector output by the last high-frequency detail restoration module after high-frequency detail restoration. This is the SAR image output by this network after speckle noise suppression. conv2 is a 3×3 convolution operation with 64 input channels and 1 output channel.

[0082] As one embodiment, the present invention provides a detailed description of the intensity domain speckle noise suppression module mentioned above, the structural schematic of which is shown below. Figure 2 As shown below, first, the high-dimensional feature vector input to this module is normalized and then subjected to a 3×3 convolution operation, as follows:

[0083] (17)

[0084] in, , These are the high-dimensional feature vectors input to this module and the high-dimensional feature vectors for initial suppression of speckle noise, respectively. LayerNorm is the layer normalization operation, and conv3 is a 3×3 convolution operation.

[0085] Next, the high-dimensional feature vector that initially suppresses speckle noise will be used. We input four consecutive convolutional layers and an activation function to perform deep speckle noise suppression, as shown below:

[0086] (18)

[0087] in, The vectors are high-dimensional feature vectors for deep suppression of speckle noise. conv4, conv5, conv6, and conv7 are convolution operations of size 1×1, 3×3, 1×1, and 3×3, respectively. sigmod is the torch.nn.sigmod() activation function in PyTorch.

[0088] Finally, a self-attention mechanism is used to deeply suppress speckle noise in high-dimensional feature vectors. After processing, it undergoes convolution and is then added to the high-dimensional feature vector input from this module. The high-dimensional feature vector after suppressing speckle noise, output by this module, is shown below:

[0089] (19)

[0090] in, This is the high-dimensional feature vector output by this module after suppressing speckle noise. AdaptiveAvgPool2d is the average pooling layer operation, and conv8 is a 1×1 convolution operation.

[0091] As one embodiment, the present invention provides a detailed description of the complex number field real and imaginary part restoration module mentioned above, the structural diagram of which is shown below. Figure 3 As shown below. First, the high-dimensional feature vector input to this module is normalized by layer and then subjected to Fast Fourier Transform to transform it from the intensity domain into a feature vector of real and imaginary coefficients in the complex domain, as shown below:

[0092] (20)

[0093] in, , These are the eigenvectors of the real part coefficients and the eigenvectors of the imaginary part coefficients in the complex field, respectively. This is the high-dimensional feature vector input to this module. LayerNorm is the layer normalization operation, and FFT is the fast Fourier transform operation.

[0094] Next, feature restoration is performed on the eigenvectors of the real and imaginary coefficients in the complex field, as shown below:

[0095] (twenty one)

[0096] (twenty two)

[0097] in, , conv9, conv10, conv11, and conv12 are the real part coefficient feature vectors and imaginary part coefficient feature vectors after feature restoration, respectively. conv9, conv10, conv11, and conv12 are all 3×3 convolution operations, and relu is the torch.nn.relu() activation function in PyTorch.

[0098] Finally, the real and imaginary coefficient eigenvectors after feature restoration are restored to the intensity domain, multiplied by the input of this module, and then added back to the input of this module to obtain the high-dimensional feature vector of the intensity domain output by this module, as shown below:

[0099] (twenty three)

[0100] in, This is the restored high-dimensional feature vector of the intensity domain output by this module.

[0101] As one embodiment, the present invention provides a detailed description of the dilated convolutional attention module mentioned above, and its structural schematic diagram is shown below. Figure 4 As shown below, the high-dimensional feature vector input to this module is first normalized and then processed through two consecutive convolutional layers to obtain preliminary restored high-frequency detail features, as shown below:

[0102] (twenty four)

[0103] in, , These are the high-dimensional feature vector input to this module and the high-dimensional vector for preliminary restoration of high-frequency detail features, respectively. LayerNorm is the layer normalization operation, and conv13 and conv14 are 1×1 and 3×3 convolution operations, respectively.

[0104] Next, the high-dimensional vectors that initially reconstruct high-frequency detail features will be used. The inputs are fed into two dilated convolutional layers of different sizes and dilation coefficients to extract feature vectors at different receptive fields and scales, as shown below:

[0105] (25)

[0106] (26)

[0107] Wherein, dconv1 and dconv2 are 3×3 dilated convolution operations with a dilation coefficient of 9 and 5×5 dilated convolution operations with a dilation coefficient of 4, respectively. , These are the feature vectors at two different receptive fields and scales extracted by dilated convolutions dconv1 and dconv2.

[0108] Then, the two feature vectors extracted above from different receptive fields at multiple scales are... , After feature concatenation, the data is fed into a dilated convolution and an activation function to restore high-frequency features, as shown below:

[0109] (27)

[0110] in, To further restore the feature vectors of high-frequency features, dconv3 is a 3×3 dilated convolution operation with a dilation coefficient of 4, relu is the torch.nn.relu() activation function in PyTorch, and cat is a feature concatenation operation that performs feature concatenation in the channel dimension.

[0111] Finally, the feature vectors for further recovery of high-frequency features... After processing using a self-attention mechanism, the input is fed into a convolutional operation and then into the input of this module, resulting in the feature vector output by this module after high-frequency feature restoration, as shown below:

[0112] (28)

[0113] in, This is the high-dimensional feature vector after high-frequency feature restoration output by this module. AdaptiveAvgPool2d is the average pooling layer operation, and conv15 is a 1×1 convolution operation.

[0114] As one embodiment, the present invention provides a detailed description of the high-frequency detail restoration module mentioned above, and its structural schematic diagram is shown below. Figure 5 As shown below, the high-dimensional feature vector input to this module is first normalized and then processed through a convolutional layer and activation function to obtain preliminary restored high-frequency detail features, as shown below:

[0115] (29)

[0116] in, , These are the high-dimensional feature vector input to this module and the feature vector for preliminary restoration of high-frequency features, respectively. LayerNorm is the layer normalization operation, conv16 is the 3×3 convolution operation, and relu is the torch.nn.relu() activation function in PyTorch.

[0117] Finally, the feature vector of the initially restored high-frequency features is input into a convolutional layer and then added to the input of this module to obtain the high-dimensional feature vector output by this module, as shown below:

[0118] (30)

[0119] in, This is the high-dimensional feature vector output by this module, and conv17 is a 1×1 convolution operation.

[0120] Please refer to the test results of this invention. Figure 6 , Figure 6 This is a schematic diagram of the noise suppression results of the method of the present invention. Figure 6 In the image, (a) represents the input SAR image with speckle noise; Figure 6 (b) in the figure represents the SAR image after speckle noise suppression using this method.

[0121] Please see Figure 7 , Figure 7 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: a SAR image speckle noise suppression device 401 based on a multi-scale strong complex domain alternating restoration network, a processor 402, and a storage device 403.

[0122] A SAR image speckle noise suppression device 401 based on a multi-scale strong complex domain alternating restoration network: The SAR image speckle noise suppression device 401 based on a multi-scale strong complex domain alternating restoration network implements the SAR image speckle noise suppression method based on a multi-scale strong complex domain alternating restoration network.

[0123] Processor 402: The processor 402 loads and executes the instructions and data in the storage device 403 to implement the SAR image speckle noise suppression method based on a multi-scale strong complex domain alternating restoration network.

[0124] Storage device 403: The storage device 403 stores instructions and data; the storage device 403 is used to implement the SAR image speckle noise suppression method based on a multi-scale strong complex domain alternating restoration network.

[0125] In summary, the beneficial effects of this invention are as follows: It proposes a SAR image speckle noise suppression method based on a multi-scale strong complex domain alternating restoration network. The main body of the network adopts a three-layer UNet structure. Its encoding part focuses on speckle noise suppression, while its decoding part focuses on high-frequency detail restoration and boundary feature sharpening. The encoding part uses alternating connected intensity domain speckle noise suppression modules and complex domain real and imaginary part restoration modules to alternately perform speckle noise suppression and complex domain restoration operations, so as to suppress speckle noise without losing the amplitude and phase information contained therein. The decoding part uses a dilated convolutional attention module and a high-frequency detail restoration module to add extra attention to the high-frequency texture part for restoration, so as to obtain a SAR image after speckle noise suppression. Finally, this invention enables high-fidelity restoration of SAR images.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A SAR image speckle noise suppression method based on a multi-scale strong complex domain alternating restoration network, characterized by: The method comprises the following steps: S1, acquiring an input SAR image with coherent speckle noise and extracting a shallow high-dimensional feature vector thereof; S2, constructing and pre-training a multi-scale strong complex domain alternating restoration network; the multi-scale strong complex domain alternating restoration network comprises an encoding part and a decoding part; S3, inputting the shallow high-dimensional feature vector into the encoding part for coherent speckle noise suppression to obtain a processed image; S4, inputting the processed image into the decoding part to obtain high-frequency detail restoration and obtain a final SAR image with coherent speckle noise suppression; The encoding part in step S2 comprises: an intensity domain coherent speckle noise suppression module, a complex domain real part and imaginary part restoration module, and a down-sampling module; Step S3 is specifically as follows: S31, inputting the shallow high-dimensional feature vector into the intensity domain coherent speckle noise suppression module to preliminarily suppress the coherent speckle noise in the intensity domain at the first scale; S32, the feature vector after suppressing the coherent speckle noise In the input complex domain real part and imaginary part recovery module, the real part and the imaginary part are respectively recovered in the complex domain, and then the intensity domain feature vector is calculated ; S33, the intensity domain feature vector after recovery , the feature vector after suppressing the speckle noise is input to the down-sampling module to continue the suppression of the speckle noise at the second scale and the recovery of the real part and the imaginary part information in the complex domain, and a feature vector after suppressing the speckle noise at the second scale is obtained , the feature vector of the intensity domain after recovery ; S34, 2 times down-sampling is performed, the above S33 step is repeated, a high-dimensional vector after the coherent speckle noise is sufficiently suppressed at a third scale is obtained, and a recovered intensity domain feature vector of an encoding part output is obtained, including a feature vector after coherent speckle noise suppression at the third scale , a feature vector of a recovered intensity domain RF 3; The decoding part in step S2 comprises: a cavity convolution attention module and a high-frequency detail restoration module; Step S4 is specifically as follows: S41, input the restored intensity domain feature vector output by the encoding part to the hollow convolution attention module and the high-frequency detail restoration module for high-frequency detail restoration and boundary feature sharpening to obtain the feature vector after attention processing at the third scale DCF 3, the feature vector after high-frequency detail restoration ; S42, the feature vector after high-frequency detail restoration under the third scale After upsampling, the feature information under the second scale is obtained, and then the restored intensity domain feature vector output by the encoding part under the second scale is added , input into the attention module and the high-frequency detail restoration module, to restore the feature information under the decoding part of the second scale, to obtain the attention processed feature vector under the second scale DCF 2, the feature vector after high-frequency detail restoration ; S43, 2 times upsampling is performed again, step S42 is repeated, and a feature vector with high frequency detail recovery and boundary feature sharpening at the first scale is obtained, including DCF 1 and ; S44, the feature vector after high-frequency detail restoration under the first scale plus the shallow high-dimensional feature vector extracted by the network initially input into the continuously connected cavity convolution attention module and high-frequency detail restoration module to perform the last step of high-frequency texture restoration work, and obtain the further processed attention processed feature vector output by the cavity convolution attention module and the high-frequency detail restoration module DCF 0, the feature vector after high-frequency detail restoration ; Step S4 further comprises: S45, the high frequency detail restored feature vector output by the last high frequency detail restoration module The input is fed into a convolutional layer, which restores it from the vector domain to the image domain, resulting in the final SAR image with speckle noise suppression.

2. A storage device, characterized by: The storage device stores instructions and data for implementing the SAR image coherent speckle noise suppression method based on the multi-scale strong complex domain alternating restoration network according to claim 1.

3. A SAR image speckle noise suppression device based on a multi-scale strong complex domain alternating restoration network, characterized by: It comprises: A processor and a storage device; the processor loads and executes the instructions and data in the storage device to implement the SAR image coherent speckle noise suppression method based on the multi-scale strong complex domain alternating restoration network according to claim 1.

Images