Intensity format SAR (Synthetic Aperture Radar) image speckle noise suppression method based on intensity complex field conversion network

Abstract

The invention relates to the field of image noise suppression, and discloses an intensity format SAR (Synthetic Aperture Radar) image speckle noise suppression method based on an intensity complex field conversion network, which comprises the following steps of: constructing the intensity complex field conversion network; training the intensity complex field conversion network by adopting a self-building governor learning mode; inputting an intensity format SAR image damaged by the speckle noise to the trained intensity complex field conversion network to obtain an image in a complex field format; constructing and pre-training a multi-receptive-field space attention speckle suppression network; inputting the image in the complex field format into the trained multi-receptive-field space attention speckle suppression network to obtain an output high-dimensional feature vector; constructing and pre-training an intensity domain detail recovery network; inputting the high-dimensional feature vector into a trained intensity domain detail restoration network to obtain a final restored SAR image in an intensity format after speckle noise suppression; according to the method, the SAR image in the intensity format damaged by the speckle noise can be restored in a high-fidelity manner.

Description

Technical Field

[0001] The present invention relates to the field of image noise suppression, and in particular to a method for suppressing speckle noise in an intensity format SAR image based on an intensity complex domain conversion network. Background Art

[0002] SAR images are widely used in various industries because they are not affected by darkness, bad weather and other conditions. SAR data commonly released by commercial satellites can generally be divided into two data formats: raw data format and processed data format. The raw data format is usually single-view complex data, stored as high-bit-width complex data, containing the amplitude and phase information of the detection unit. It cannot be displayed in the form of an image, and the data is large and difficult to transmit and store. The processed data format is usually low-bit-width 8-bit intensity format data, which can be displayed in the form of an image. The data is small and easy to transmit and store, but it loses a lot of radiation and phase information due to the 8-bit bit width. Therefore, SAR intensity format data is usually released, disseminated, and used, and a large amount of application work is carried out based on intensity format data.

[0003] However, due to the limitations of the coherent imaging principle, SAR images contain a large amount of bright and dark fluctuating coherent speckle noise in the acquired data, which seriously affects the subsequent interpretation and related applications. In order to solve the related problems caused by coherent speckle noise, researchers have proposed a large number of research methods, which can be mainly divided into coherent speckle noise suppression methods based on spatial domain, transform domain, variational method, and deep learning, showing good coherent speckle noise suppression results. However, most of their work is based on SAR images in intensity format. The process of converting complex data into intensity format naturally loses a lot of information such as radiation, phase, and data accuracy. It is difficult to restore SAR images damaged by coherent speckle noise with high fidelity based on radiation characteristics. Summary of the Invention

[0004] The purpose of the present invention is to propose a method for suppressing speckle noise in intensity-format SAR images based on an intensity-complex-domain conversion network, so as to solve the problem that existing intensity-format SAR images damaged by speckle noise are difficult to restore with high fidelity in the intensity domain.

[0005] Specifically, the present invention provides a method for suppressing speckle noise in intensity format SAR images based on an intensity complex domain conversion network, comprising the following steps: S1. Construct the intensity complex domain conversion network ICCNet; S2. Use self-built supervised learning to train the intensity complex domain conversion network ICCNet to obtain a trained intensity complex domain conversion network; S3, inputting the intensity-format SAR image corrupted by speckle noise into the trained intensity-complex-domain conversion network to obtain a complex-domain SAR image; S4. Construct and pre-train a multi-receptive field spatial attention speckle suppression network; S5. Inputting the SAR image in the complex domain format into the trained multi-receptive field spatial attention coherent speckle suppression network to obtain a high-dimensional feature vector output by the multi-receptive field spatial attention coherent speckle suppression network; S6. construct and pre-train an intensity domain detail restoration network to obtain a trained intensity domain detail restoration network; S7. Input the high-dimensional feature vector output by the multi-receptive field spatial attention coherent speckle suppression network into the trained intensity domain detail restoration network to obtain the final restored SAR image in the intensity format after coherent speckle noise suppression.

[0006] A storage device stores instructions and data for implementing an intensity format SAR image speckle noise suppression method based on an intensity complex domain conversion network.

[0007] A device for suppressing speckle noise in an intensity-format SAR image based on an intensity-complex domain conversion network comprises: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a method for suppressing speckle noise in an intensity-format SAR image based on an intensity-complex domain conversion network.

[0008] The present invention provides the following beneficial effects: a method for suppressing speckle noise in an intensity-format SAR image based on an intensity-complex domain conversion network is proposed, wherein the intensity-format SAR image is converted into a complex-domain SAR image through the intensity-complex domain conversion network, and then the real coefficients and imaginary coefficients converted into the complex domain are distributed and input into a multi-receptive field spatial attention speckle suppression network to complete speckle noise suppression of the corresponding complex-domain real coefficient vectors and imaginary coefficient vectors. Subsequently, the intensity domain vector is calculated using the obtained real coefficient vector and imaginary coefficient vector, and then input into an intensity domain detail restoration network to restore some high-frequency details that cannot be restored by the speckle suppression network, thereby obtaining the final intensity-format SAR image after speckle noise suppression. In general, the present invention can restore the intensity-format SAR image damaged by speckle noise with high fidelity. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a flow chart of the method for suppressing speckle noise in intensity format SAR images based on an intensity complex domain conversion network according to the present invention; Figure 2 This is a schematic diagram of the intensity complex domain conversion network ICCNet; Figure 3It is a structural diagram of densely connected modules; Figure 4 This is a schematic diagram of a multi-receptive field spatial attention speckle suppression network; Figure 5 This is a schematic diagram of the multi-receptive field spatial attention module; Figure 6 is a schematic diagram of the intensity domain detail restoration network; Figure 7 It is a schematic diagram of the results of the method of the present invention; Figure 8 This is a working diagram of the hardware equipment for this application. DETAILED DESCRIPTION

[0010] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0011] Before formally explaining the present invention, the scheme of the present invention is first generally explained for easy understanding.

[0012] Please refer to Figure 1 The present invention provides a method for suppressing speckle noise in intensity format SAR images based on an intensity complex domain conversion network, comprising the following steps: S1. Construct the intensity complex domain conversion network ICCNet; S2. Use self-built supervised learning to train the intensity complex domain conversion network ICCNet to obtain a trained intensity complex domain conversion network; S3, inputting the intensity-format SAR image corrupted by speckle noise into the trained intensity-complex-domain conversion network to obtain a complex-domain SAR image; S4. Construct and pre-train a multi-receptive field spatial attention speckle suppression network; S5. Inputting the SAR image in the complex domain format into the trained multi-receptive field spatial attention coherent speckle suppression network to obtain a high-dimensional feature vector output by the multi-receptive field spatial attention coherent speckle suppression network; S6. construct and pre-train an intensity domain detail restoration network to obtain a trained intensity domain detail restoration network; S7. Input the high-dimensional feature vector output by the multi-receptive field spatial attention coherent speckle suppression network into the trained intensity domain detail restoration network to obtain the final restored SAR image in the intensity format after coherent speckle noise suppression.

[0013] It should be noted that in step S1, the intensity complex domain conversion network ICCNet includes: a shallow feature extraction module, a first intensity domain information encoding module, a second intensity domain information encoding module, a first real coefficient restoration module, a second real coefficient restoration module, a first imaginary coefficient restoration module, a second imaginary coefficient restoration module and a convolutional layer arranged in sequence.

[0014] It should be noted that the first intensity domain information encoding module, the second intensity domain information encoding module, the first real coefficient restoration module, the second real coefficient restoration module, the first imaginary coefficient restoration module, and the second imaginary coefficient restoration module have the same structure, and all include: 3 groups of alternately connected 3×3 convolutional layers arranged in sequence, activation functions, a feature splicing module, and a 1×1 convolutional layer.

[0015] It should be noted that when self-supervised learning is used in step S2, the loss function is as follows:

[0016] in, is the L1 norm, They are SAR images in real complex format The real coefficient and imaginary coefficient of ; They are respectively the complex format SAR images converted by the intensity complex domain conversion network The real and imaginary coefficients of .

[0017] It should be noted that step S3 is specifically as follows: S31. Intensity format SAR image corrupted by speckle noise Input to the shallow feature extraction module to obtain the shallow feature vector ; S32, shallow feature vector Input into the first intensity domain information encoding module and the second intensity domain information encoding module in sequence, and obtain the first high-dimensional feature vector encoded by the intensity domain information encoding module ; S33, the first high-dimensional feature vector encoded by the intensity domain information encoding module Perform Fourier transform to obtain the first real coefficient eigenvector in the complex domain after transformation and the second imaginary coefficient eigenvector ; S34, the first real coefficient eigenvector Input to the first real coefficient recovery module and the second real coefficient recovery module in sequence, and convert the second imaginary coefficient eigenvector Input to the first imaginary coefficient recovery module and the second imaginary coefficient recovery module connected in sequence to obtain the second real coefficient high-dimensional feature vector , the second imaginary coefficient high-dimensional eigenvector ; S35, the second real part coefficient high dimensional eigenvector , the second imaginary coefficient high-dimensional eigenvector Input them into the convolution layer respectively to obtain the real coefficients of the intensity complex domain conversion network output , imaginary coefficient ; S36, real coefficients output by the intensity complex domain conversion network , imaginary coefficient SAR images in complex domain format.

[0018] It should be noted that the multi-receptive field spatial attention speckle suppression network in step S4 includes: a real multi-receptive field spatial attention speckle suppression network SMSTNet and an imaginary multi-receptive field spatial attention speckle suppression network XMSTNet; wherein XMSTNet and XMSTNet have the same structure, both including: a convolutional layer, a first multi-receptive field spatial attention module, a second multi-receptive field spatial attention module, a third multi-receptive field spatial attention module, a feature splicing module, and a convolutional layer arranged in sequence.

[0019] It should be noted that step S5 is specifically as follows: Step S5 is specifically as follows: S51, the second real part coefficient high dimensional eigenvector , the second imaginary coefficient high-dimensional eigenvector F b2 Input them into 1 convolution layer respectively for feature vector extraction to obtain the high-dimensional feature vector of the third real part coefficient , the third imaginary coefficient high-dimensional eigenvector F b3 ; S52, the third real part coefficient high dimensional eigenvector Input into the first multi-receptive field spatial attention module, the second multi-receptive field spatial attention module, and the third multi-receptive field spatial attention module connected in sequence to obtain the fourth real part coefficient high-dimensional feature vector , the fifth real part coefficient high-dimensional eigenvector , the sixth real part coefficient high-dimensional eigenvector ; S52, the third imaginary coefficient high-dimensional eigenvector F b3 Input into the first multi-receptive field spatial attention module, the second multi-receptive field spatial attention module, and the third multi-receptive field spatial attention module connected in sequence to obtain the fourth imaginary part coefficient high-dimensional feature vector F b4, the fifth imaginary coefficient high-dimensional eigenvector F b5 , the sixth imaginary coefficient high-dimensional eigenvector F b6 ;; S53, the fourth real part coefficient high dimensional eigenvector , the fifth real part coefficient high-dimensional eigenvector , the sixth real part coefficient high-dimensional eigenvector After being spliced ​​by the feature splicing module, it is convolved by the convolution layer and then combined with the second real part coefficient high-dimensional feature vector Add together to get the real high-dimensional feature vector output by the real multi-receptive field spatial attention speckle suppression network SMSTNet ; S54, the fourth imaginary coefficient high-dimensional eigenvector F b4 , the fifth imaginary coefficient high-dimensional eigenvector F b5 , the sixth imaginary coefficient high-dimensional eigenvector F b6 After being spliced ​​by the feature splicing module, it is convolved by the convolution layer and then combined with the second imaginary coefficient high-dimensional feature vector F b2 Add them together to get the imaginary high-dimensional feature vector F output by the imaginary multi-receptive field spatial attention coherent speckle suppression network XMSTNet bd ; S55, the real part of the multi-receptive field spatial attention speckle suppression network SMSTNet output real high-dimensional feature vector The imaginary high-dimensional feature vector F output by the multi-receptive field spatial attention coherent speckle suppression network XMSTNet bd Convert to the intensity domain to obtain the high-dimensional feature vector output by the multi-receptive field spatial attention coherent speckle suppression network .

[0020] It should be noted that the intensity domain detail restoration network in step S6 includes: a first multi-receptive field spatial attention module, a second multi-receptive field spatial attention module, a third multi-receptive field spatial attention module, a feature splicing module, and a convolutional layer arranged in sequence; The high-dimensional feature vector output by the multi-receptive field spatial attention speckle suppression network The image is input into the intensity domain detail restoration network to obtain the final restored SAR image in intensity format after speckle noise suppression.

[0021] As an embodiment, the present invention further elaborates on the intermediate process of the method.

[0022] First, the present invention inputs the intensity-format SAR image corrupted by coherent speckle noise into the intensity-complex domain conversion network, converts the intensity-format SAR image from the intensity domain to the complex domain, and expresses the corresponding radiation and phase information in the complex domain with high precision and high bit width. The specific process is as follows: (1) in, They are complex domain and intensity domain SAR images respectively. ICCNet is the intensity complex domain conversion network proposed in this invention. The corresponding network structure diagram is shown in Figure 2 In addition, the complex domain SAR image can be represented by the real coefficients and the imaginary coefficient To express, such as , where i is an imaginary unit, the subsequent steps of the present invention are respectively and the imaginary coefficient Suppressing coherent speckle noise.

[0023] Next, the present invention elaborates on the specific steps of the intensity complex domain conversion network ICCNet. For ICCNet, the present invention adopts a self-built supervised learning method to train it. After the training is completed, its model parameters are frozen. Only the inference process is performed in the intensity format SAR image coherent speckle noise suppression network based on the intensity complex domain conversion network. The structural diagram of the intensity format SAR image coherent speckle noise suppression network based on the intensity complex domain conversion network is shown in FIG. Figure 1 shown.

[0024] The present invention uses SAR images in real complex format To calculate the SAR image in intensity format , then Input to the intensity complex domain conversion network ICCNet to obtain the converted complex format SAR image , constructed by The data pairs are used to establish L1 loss for self-supervised training of ICCNet, as shown below: (2) (3) (4) in, They are respectively the real complex format SAR image, the calculated real intensity format SAR image, and the complex format SAR image converted by the intensity complex domain conversion network. is a SAR image in real complex format The real coefficient and imaginary coefficient of the intensity complex domain conversion network. Complex format SAR image after conversion The real and imaginary coefficients of To express it, the established self-supervised loss function is as follows: (5) Among them, LOSS_{ICCNet} is the self-supervised loss function of the training intensity complex domain conversion network, is the L1 norm. The SAR image in the true intensity format obtained by calculation , the present invention inputs it into the shallow feature extraction module and converts it into a shallow feature vector as shown below: (6) in, The shallow feature vector is extracted by the shallow feature extraction module. Conv1 is a 3×3 convolution with 1 input channel and 64 output channels.

[0025] Then, the present invention converts the shallow feature vector The intensity image is sequentially input into the first intensity domain information encoding module and the second intensity domain information encoding module to fully represent the radiation and phase information contained in the intensity image, as shown below: (7) in, The first high-dimensional feature vector after encoding by the intensity domain information encoding module, INCM1, INCM2 are respectively the first intensity domain information encoding module and the second intensity domain information encoding module. In addition, the two intensity domain information encoding modules described above adopt the same structure, which is a densely connected module structure. Figure 3 Therefore, the present invention will be described in detail below based on the first intensity domain information encoding module.

[0026] The present invention converts shallow feature vector First, it is fed into a convolutional layer and activation function as follows: (8) in, is the encoded feature vector, relu is the torch.nn.Relu function in pytorch, conv2 is a 3×3 convolution, and both the input and output channels are 64. Next, the present invention repeats the above operation twice, as shown below: (9) (10) in, Both are fully encoded feature vectors, conv3 and conv4 are 3×3 convolutions, and the input and output channels are both 64.

[0027] Then, the present invention calculates the encoded feature vectors of the above steps successively and the shallow feature vector of the input Perform splicing as follows: (11) in, is the high-dimensional feature vector after feature splicing, and cat is the operation of feature splicing in the channel dimension.

[0028] The present invention then combines the high-dimensional feature vector Input to the convolution layer, and then add the shallow feature vector input by this module , and get the output of this module as follows: (12) in, This is the high-dimensional feature vector after encoding the intensity domain information output by this module. Conv5 is a 1×1 convolution operation with 256 input channels and 64 output channels.

[0029] Through the above steps, the present invention obtains the high-dimensional feature vector after encoding by the first intensity domain information encoding module and the second intensity domain information encoding module. , then the present invention performs Fourier transform on the alignment to convert it into real coefficient eigenvector and imaginary coefficient eigenvector, as shown below: (13) in, are the first real coefficient eigenvector and the first imaginary coefficient eigenvector converted into the complex domain respectively, and FFT is fast Fourier transform.

[0030] Then, the present invention inputs the real coefficient eigenvector and the imaginary coefficient eigenvector into two consecutive first real coefficient recovery modules, second real coefficient recovery modules and first imaginary coefficient recovery modules, second imaginary coefficient recovery modules, respectively, as shown below: (14) (15) in, They are respectively the restored high-dimensional eigenvector of the second real coefficient and the high-dimensional eigenvector of the second imaginary coefficient, SCRB1 and SCRB2 are respectively the first real coefficient restoration module and the second real coefficient restoration module, XCRB1 and XCRB2 are respectively the first imaginary coefficient restoration module and the second imaginary coefficient restoration module. The network structure adopted by the first real coefficient restoration module, the second real coefficient restoration module, the first imaginary coefficient restoration module and the second imaginary coefficient restoration module is the same as that of the intensity domain information encoding module, which is a densely connected module structure, so it will not be described in detail. The corresponding schematic diagram is as follows: Figure 3 shown.

[0031] Finally, the present invention converts the restored real coefficient high-dimensional eigenvector , high-dimensional eigenvector of imaginary coefficient Input to two convolutional layers to obtain the real coefficients of the intensity complex domain conversion network output , imaginary coefficient , as shown below: (16) (17) in, They are the real coefficient and imaginary coefficient of the intensity complex domain conversion network output, respectively. Conv6 and Conv7 are both 3×3 convolution layers with 64 input channels and 1 output channel.

[0032] Through the intensity complex domain conversion network, the present invention converts the intensity format SAR image from the intensity domain to the complex domain, and can use high-dimensional complex domain vectors to express radiation and phase information with high precision and redundancy. Next, the present invention uses the multi-receptive field spatial attention coherent speckle suppression network in the complex domain to perform high-dimensional feature vectors of real coefficients. , high-dimensional eigenvector of imaginary coefficient To suppress the speckle noise contained therein. Specifically, the present invention inputs the second real coefficient high-dimensional eigenvector and the second imaginary coefficient high-dimensional eigenvector into the real multi-receptive field spatial attention coherent speckle suppression network and the imaginary multi-receptive field spatial attention coherent speckle suppression network, respectively, as shown below: (18) (19) in, are respectively the high-dimensional eigenvector of the real coefficient and the high-dimensional eigenvector of the imaginary coefficient after the coherent speckle noise is suppressed. They are respectively the real coefficient high-dimensional feature vector and the imaginary coefficient high-dimensional feature vector after the intensity complex domain conversion network, SMSTNet and XMSTNet are respectively the real multi-receptive field spatial attention coherent speckle suppression network and the imaginary multi-receptive field spatial attention coherent speckle suppression network. Their network structures are the same, but they serve as two branches to process the coherent speckle noise in the real and imaginary domains respectively. Therefore, in this article, the present invention only elaborates on the real multi-receptive field spatial attention coherent speckle suppression network. Its network structure is as follows Figure 4 shown.

[0033] First, the present invention converts the real coefficient high-dimensional eigenvector Input into a convolutional layer for preliminary feature vector extraction, as shown below: (20) in, It is the deep feature vector after convolution operation. Conv8 is a 3×3 convolution operation with 64 input and output channels. Next, the present invention processes the deep feature vector Three multi-receptive field spatial attention modules with continuous input values ​​are used to fully remove the coherent speckle noise contained in the real coefficient vector, as shown below: (twenty one) (twenty two) (twenty three) Among them, MSTBlock1, MSTBlock2, and MSTBlock3 are the first multi-receptive field spatial attention module, the second multi-receptive field spatial attention module, and the third multi-receptive field spatial attention module, respectively. are the high-dimensional feature vectors after the corresponding spatial attention modules MSTBlock1, MSTBlock2, and MSTBlock3 suppress the coherent speckle noise. The network structures used by MSTBlock1, MSTBlock2, and MSTBlock3 are the same, and the corresponding structural diagrams are shown in the figure below. Figure 5 As shown, the present invention is described in detail only with respect to MSTBlock1. The input is fed into two consecutive convolutional layers as follows: (twenty four) in, For the high-dimensional feature vectors that are preliminarily processed, conv9 is a 1×1 convolution operation, and conv10 is a 3×3 convolution operation. Both their input and output channels are 64. Then, the present invention inputs F_{as1} into three convolution operations of different scales to learn features of different receptive fields, and then performs feature splicing, as shown below: (25) in, is the 192-dimensional high-dimensional feature vector after feature splicing, cat is the feature splicing operation in the channel dimension, conv11, conv12, and conv13 are 3×3, 5×5, and 7×7 convolution operations respectively, and their input channels and output channels are all 64. Next, the present invention inputs the obtained 192-dimensional high-dimensional feature vector into a 3×3 convolutional layer to fully fuse the spatial information of multiple receptive fields, as shown below: (26) in, It is a high-dimensional feature vector after the spatial information of multiple receptive fields is fused. Conv14 is a 3×3 convolution operation with 192 input channels and 64 output channels. Then, the present invention uses activation function to Generate self-attention weights to weight the self and add them as the residual to the input, as shown below: (27) in, is the high-dimensional feature vector processed by the self-attention mechanism, and sigmod is the torch.nn.sigmod activation function in pytorch. Finally, the present invention will After inputting the 1×1 convolutional layer and adding the input of this module, we get the feature vector output by MSTBlock1 of this module, as shown below: (28) in, They are the high-dimensional feature vectors of the input and output of MSTBlock1 of this module respectively. conv15 is a 1×1 convolution operation, and both the input and output channels are 64. Through the above process, the present invention has developed a detailed description of the first multi-receptive field spatial attention module MSTBlock1 and described the corresponding network structure process. The present invention obtains the high-dimensional feature vectors of the corresponding modules after the coherent speckle noise is suppressed by arranging three multi-receptive field spatial attention modules in sequence. , after splicing it in the feature dimension, it is input into the final convolutional layer and then added with the input of the real multi-receptive field spatial attention coherent speckle suppression network , get the real high-dimensional feature vector output by the network: (29) in, are the high-dimensional feature vectors of the real coefficients before and after speckle noise suppression, cat is the feature concatenation operation in the channel dimension, conv16 is a 3×3 convolution operation, the input channels are 192, and the output channels are both 64.

[0034] Through the real part multi-receptive field spatial attention coherent speckle suppression network SMSTNet and the imaginary part multi-receptive field spatial attention coherent speckle suppression network XMSTNet, the present invention obtains the real part coefficient high-dimensional feature vector after coherent speckle noise suppression , high-dimensional eigenvector of imaginary coefficient Next, the present invention transforms the real and imaginary high-dimensional feature vectors in the complex domain into the intensity domain, laying the foundation for the subsequent recovery of high-frequency details in the intensity domain, as shown below: (30) in, It converts the real and imaginary high-dimensional feature vectors in the complex domain into high-dimensional feature vectors in the intensity domain.

[0035] Through the above steps, the present invention obtains the high-dimensional feature vector of the intensity domain after the coherent speckle noise is suppressed. , then the present invention inputs it into the intensity domain detail restoration network to restore some high-frequency details that the coherent speckle suppression network cannot restore. The intensity domain detail restoration network includes three multi-receptive field spatial attention modules, a feature splicing operation and a convolution operation. The corresponding structural diagram is shown in the figure. Figure 6 As shown in the figure, the three multi-receptive field spatial attention modules used are the same as those used in the multi-receptive field spatial attention speckle suppression network in step (2). They have been described in detail in step (2), so they will not be described in detail here. The structural diagram of the corresponding modules is shown in the figure. Figure 5 shown.

[0036] For the intensity domain detail restoration network, the present invention sequentially inputs the high-dimensional feature vector of the intensity domain after speckle noise suppression into three multi-receptive field spatial attention modules to restore some high-frequency details in the intensity domain, as shown below: (31) (32) (33) Among them, MSTBlock4, MSTBlock5, and MSTBlock6 are multi-receptive field spatial attention modules 4, 5, and 6 respectively. They are the high-dimensional feature vectors after the high-frequency details of the corresponding spatial attention modules MSTBlock4, MSTBlock5, and MSTBlock6 are restored. The present invention obtains the high-dimensional feature vector after the high-frequency details of the corresponding spatial attention module are restored by arranging three multi-receptive field spatial attention modules in sequence. , and then it is spliced ​​in the feature dimension and input into the final convolution layer to obtain the SAR image in intensity format after the coherent speckle noise suppression output by this method, as shown below: (34) Where out is the SAR image in intensity format after speckle noise suppression output by this method, cat is the feature concatenation operation in the channel dimension, conv17 is a 3×3 convolution operation, the input channels are 192, and the output channels are both 64.

[0037] Please refer to Figure 7 , Figure 7 Schematic diagram of the results of the method of the present invention; wherein Figure 7 (a) in the figure represents the input SAR image with coherent speckle noise; Figure 7 (b) in the figure shows the SAR image after the coherent speckle noise is suppressed by the method of the present invention.

[0038] See Figure 8 , Figure 8 4 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: an intensity format SAR image speckle noise suppression device 401 based on an intensity complex domain conversion network, a processor 402 and a storage device 403.

[0039] An intensity format SAR image speckle noise suppression device 401 based on an intensity complex domain conversion network: The intensity format SAR image speckle noise suppression device 401 based on an intensity complex domain conversion network implements the intensity format SAR image speckle noise suppression method based on an intensity complex domain conversion network.

[0040] Processor 402: The processor 402 loads and executes instructions and data in the storage device 403 to implement the intensity format SAR image speckle noise suppression method based on the intensity complex domain conversion network.

[0041] Storage device 403: The storage device 403 stores instructions and data; the storage device 403 is used to implement the intensity format SAR image speckle noise suppression method based on the intensity complex domain conversion network.

[0042] In general, the beneficial effects of the present invention are as follows: a method for suppressing speckle noise in intensity-format SAR images based on an intensity-complex domain conversion network is proposed, wherein the intensity-format SAR image is converted into a complex-domain SAR image through the intensity-complex domain conversion network, and then the real coefficients and imaginary coefficients converted into the complex domain are distributed and input into a multi-receptive field spatial attention speckle suppression network to complete the speckle noise suppression of the corresponding complex-domain real coefficient vectors and imaginary coefficient vectors. Subsequently, the intensity domain vector is calculated using the obtained real coefficient vector and imaginary coefficient vector and then input into an intensity domain detail restoration network to restore some high-frequency details that cannot be restored by the speckle suppression network, thereby obtaining the final intensity-format SAR image after speckle noise suppression; generally speaking, the present invention can restore the intensity-format SAR image damaged by speckle noise with high fidelity.

[0043] 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 in the scope of protection of the present invention.

Claims

1. A method for suppressing speckle noise in intensity-format SAR images based on an intensity-complex domain conversion network, characterized by: The method comprises the following steps: S1. Construct the intensity complex domain conversion network ICCNet; S2. Use self-built supervised learning to train the intensity complex domain conversion network ICCNet to obtain a trained intensity complex domain conversion network; S3, inputting the intensity-format SAR image corrupted by speckle noise into the trained intensity-complex-domain conversion network to obtain a complex-domain SAR image; S4. Construct and pre-train a multi-receptive field spatial attention speckle suppression network; S5. Inputting the SAR image in the complex domain format into the trained multi-receptive field spatial attention coherent speckle suppression network to obtain a high-dimensional feature vector output by the multi-receptive field spatial attention coherent speckle suppression network; S6. construct and pre-train an intensity domain detail restoration network to obtain a trained intensity domain detail restoration network; S7. Input the high-dimensional feature vector output by the multi-receptive field spatial attention coherent speckle suppression network into the trained intensity domain detail restoration network to obtain the final restored SAR image in the intensity format after coherent speckle noise suppression.

2. The method for suppressing speckle noise in intensity-format SAR images based on an intensity-complex domain conversion network according to claim 1, wherein: In step S1, the intensity complex domain conversion network ICCNet includes: a shallow feature extraction module, a first intensity domain information encoding module, a second intensity domain information encoding module, a first real coefficient restoration module, a second real coefficient restoration module, a first imaginary coefficient restoration module, a second imaginary coefficient restoration module and a convolutional layer arranged in sequence.

3. The method for suppressing speckle noise in intensity-format SAR images based on an intensity-complex domain conversion network according to claim 2, wherein: The first intensity domain information encoding module, the second intensity domain information encoding module, the first real coefficient restoration module, the second real coefficient restoration module, the first imaginary coefficient restoration module, and the second imaginary coefficient restoration module have the same structure, and all include: 3 groups of alternately connected 3×3 convolutional layers arranged in sequence, an activation function, a feature splicing module, and a 1×1 convolutional layer.

4. The method for suppressing speckle noise in intensity-format SAR images based on an intensity-complex domain conversion network according to claim 1, wherein: When self-supervised learning is used in step S2, the loss function is as follows: in, is the L1 norm, They are SAR images in real complex format The real coefficient and imaginary coefficient of ; They are respectively the complex format SAR images converted by the intensity complex domain conversion network The real and imaginary coefficients of .

5. The method for suppressing speckle noise in intensity-format SAR images based on an intensity-complex domain conversion network according to claim 2, wherein: Step S3 is as follows: S31. Intensity format SAR image corrupted by speckle noise Input to the shallow feature extraction module to obtain the shallow feature vector ; S32, shallow feature vector Input into the first intensity domain information encoding module and the second intensity domain information encoding module in sequence, and obtain the first high-dimensional feature vector encoded by the intensity domain information encoding module ; S33, the first high-dimensional feature vector encoded by the intensity domain information encoding module Perform Fourier transform to obtain the first real coefficient eigenvector in the complex domain after transformation and the second imaginary coefficient eigenvector ; S34, the first real coefficient eigenvector Input to the first real coefficient recovery module and the second real coefficient recovery module in sequence, and convert the second imaginary coefficient eigenvector Input to the first imaginary coefficient recovery module and the second imaginary coefficient recovery module connected in sequence to obtain the second real coefficient high-dimensional feature vector , the second imaginary coefficient high-dimensional eigenvector ; S35, the second real part coefficient high dimensional eigenvector , the second imaginary coefficient high-dimensional eigenvector Input them into the convolution layer respectively to obtain the real coefficients of the intensity complex domain conversion network output , imaginary coefficient ; S36, real coefficients output by the intensity complex domain conversion network , imaginary coefficient SAR images in complex domain format.

6. The method for suppressing speckle noise in intensity-format SAR images based on an intensity-complex domain conversion network according to claim 1, wherein: The multi-receptive field spatial attention speckle suppression network in step S4 includes: a real multi-receptive field spatial attention speckle suppression network SMSTNet and an imaginary multi-receptive field spatial attention speckle suppression network XMSTNet; wherein XMSTNet and XMSTNet have the same structure, both including: a convolutional layer, a first multi-receptive field spatial attention module, a second multi-receptive field spatial attention module, a third multi-receptive field spatial attention module, a feature splicing module, and a convolutional layer arranged in sequence.

7. The method for suppressing speckle noise in intensity-format SAR images based on an intensity-complex domain conversion network according to claim 6, wherein: Step S5 is specifically as follows: S51, the second real part coefficient high dimensional eigenvector , the second imaginary coefficient high-dimensional eigenvector F b2 Input them into 1 convolution layer respectively for feature vector extraction to obtain the high-dimensional feature vector of the third real part coefficient , the third imaginary coefficient high-dimensional eigenvector F b3 ; S52, the third real part coefficient high dimensional eigenvector Input into the first multi-receptive field spatial attention module, the second multi-receptive field spatial attention module, and the third multi-receptive field spatial attention module connected in sequence to obtain the fourth real part coefficient high-dimensional feature vector , the fifth real part coefficient high-dimensional eigenvector , the sixth real part coefficient high-dimensional eigenvector ; S52, the third imaginary coefficient high-dimensional eigenvector F b3 Input into the first multi-receptive field spatial attention module, the second multi-receptive field spatial attention module, and the third multi-receptive field spatial attention module connected in sequence to obtain the fourth imaginary part coefficient high-dimensional feature vector F b4 , the fifth imaginary coefficient high-dimensional eigenvector F b5 , the sixth imaginary coefficient high-dimensional eigenvector F b6 ;; S53, the fourth real part coefficient high dimensional eigenvector , the fifth real part coefficient high-dimensional eigenvector , the sixth real part coefficient high-dimensional eigenvector After being spliced ​​by the feature splicing module, it is convolved by the convolution layer and then combined with the second real part coefficient high-dimensional feature vector Add together to get the real high-dimensional feature vector output by the real multi-receptive field spatial attention speckle suppression network SMSTNet ; S54, the fourth imaginary coefficient high-dimensional eigenvector F b4 , the fifth imaginary coefficient high-dimensional eigenvector F b5 , the sixth imaginary coefficient high-dimensional eigenvector F b6 After being spliced ​​by the feature splicing module, it is convolved by the convolution layer and then combined with the second imaginary coefficient high-dimensional feature vector F b2 Add them together to get the imaginary high-dimensional feature vector F output by the imaginary multi-receptive field spatial attention coherent speckle suppression network XMSTNet bd ; S55, the real part of the multi-receptive field spatial attention speckle suppression network SMSTNet output real high-dimensional feature vector The imaginary high-dimensional feature vector F output by the multi-receptive field spatial attention coherent speckle suppression network XMSTNet bd Convert to the intensity domain to obtain the high-dimensional feature vector output by the multi-receptive field spatial attention coherent speckle suppression network .

8. The method for suppressing speckle noise in intensity-format SAR images based on an intensity-complex domain conversion network according to claim 7, wherein: The intensity domain detail restoration network in step S6 includes: a first multi-receptive field spatial attention module, a second multi-receptive field spatial attention module, a third multi-receptive field spatial attention module, a feature splicing module, and a convolutional layer arranged in sequence; The high-dimensional feature vector output by the multi-receptive field spatial attention speckle suppression network The image is input into the intensity domain detail restoration network to obtain the final restored SAR image in intensity format after speckle noise suppression.

9. A storage device, characterized in that: The storage device stores instructions and data for implementing the intensity format SAR image speckle noise suppression method based on an intensity complex domain conversion network as described in any one of claims 1 to 8.

10. A device for suppressing speckle noise in intensity-format SAR images based on an intensity-complex domain conversion network, characterized by: include: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the intensity format SAR image coherent speckle noise suppression method based on an intensity complex domain conversion network as described in any one of claims 1 to 8.

Images