A SAR image speckle suppression method combining network filtering and structure restoration

Abstract

The application discloses a SAR image speckle suppression method combining network filtering and structure restoration, wherein the distribution of speckles in a SAR image is related to the surface characteristics of an imaging area, and target structures such as textures, edges and strong scatterers of a heterogeneous area are often blurred or damaged in the speckle suppression process. To solve the problem, firstly, a bilateral residual block network is used to filter the SAR image to suppress speckles as much as possible, and then the damaged target structures in the filtering process are restored through image segmentation based on an edge detection network and weighted fusion based on linear weighting. Results show that the method realizes suppression of speckles while also completely preserving target structures in the SAR image.

Description

Technical Field

[0001] This invention relates to a method for processing SAR images, and more particularly to a method for suppressing speckle in SAR (Synthetic Aperture Radar) images by combining network filtering and structure restoration. Background Technology

[0002] The book *Radar Imaging Technology*, published by Harbin Institute of Technology Press in October 1999 and edited by Liu Yongtan, states that Synthetic Aperture Radar (SAR) is mounted on a moving platform and transmits and receives pulses at a certain repetition frequency to form echo signals. The structural block diagram of a SAR system is shown below. Figure 1 As shown, the SAR system consists of three parts: an onboard radar system, a satellite platform and data downlink system, and a ground system. It processes synthetic aperture radar imaging to obtain SAR images; the SAR images are stored in a backup operating system.

[0003] Synthetic Aperture Radar (SAR) is an active coherent side-looking imaging system operating in the microwave band. The radar transmits phase-modulated pulses at an angle to the sensor's motion vector, receives and records the reflected echoes, and then processes them using imaging algorithms to obtain a two-dimensional radar image. SAR systems have advantages such as long observation range and high resolution. They can also overcome limitations imposed by sunlight and weather conditions, penetrating clouds, vegetation, water surfaces, and other obstructions to achieve all-weather, 24 / 7 Earth monitoring. Therefore, they are widely used in environmental monitoring, resource exploration, and maritime vessel identification. However, the coherent imaging principle inevitably leads to speckle. When the radar beam illuminates a rough surface, the echo received by the SAR system is a coherent superposition of multiple randomly phased echoes within the resolvable cell. The amplitude of these echoes fluctuates greatly around the scattering coefficient value, resulting in speckles of randomly varying brightness in the image. The presence of speckle severely affects the visual quality of SAR images, hinders image interpretation, and significantly increases the difficulty of subsequent target detection and classification. Therefore, suppressing speckle and improving image quality are essential steps in the vast majority of SAR image applications.

[0004] The distribution of speckle is closely related to the surface roughness and backscattering characteristics illuminated by the radar beam, and different types of imaging surfaces place different requirements on the algorithm. For homogeneous regions with uniform surface characteristics, the algorithm needs to suppress speckle as much as possible; however, for heterogeneous regions with non-uniform surface characteristics, in practical applications, people often pay more attention to target structures outside the background, such as narrow edges of roads and rivers, or strong man-made scatterers such as airplanes, ships, and buildings. Therefore, while suppressing speckle, it is also necessary to keep the target structure such as texture, edges, and strong scatterers unaffected. Existing speckle suppression algorithms are mainly divided into traditional filtering methods and deep learning-based algorithms, each with its own advantages and disadvantages in terms of speckle suppression and structure preservation. Classical methods are mainly divided into local spatial domain filtering, transform domain filtering, and nonlocal filtering. Summary of the Invention

[0005] To reduce speckle interference in SAR images output by SAR processors, this invention proposes a SAR image speckle suppression method that combines network filtering and structure restoration. This method uses a bilateral residual block network for speckle filtering, and employs image segmentation based on an edge detection network and image fusion based on linear weighting to achieve speckle suppression in SAR images while preserving the target structure.

[0006] The present invention provides a SAR image speckle suppression method combining network filtering and structure restoration, which performs speckle suppression on a single-look complex image output by an 1A-level processor in a SAR signal processor; it includes the following steps:

[0007] Step 1: SAR image speckle filtering;

[0008] Step 11, Acquisition of SAR Images;

[0009] The single-look complex image output by the 1A-level processor in the SAR signal processor is denoted as the single-look complex SAR image X. in ;

[0010] Step 12: Obtain the loss function for the SAR image based on the bilateral residual block;

[0011] The network structure based on bilateral residual blocks contains two ordinary convolutional layers in both the input and output parts. Except for the final output layer, the other three layers all have batch normalization layers added before the ReLU activation function layer. The core of the network is the six cascaded bilateral residual blocks in the middle.

[0012] During training, the SAR image dataset is divided into a SAR image training set and a SAR image validation set. The SAR image training set is denoted as... The SAR image verification set is denoted as Input the data into a network based on bilateral residual blocks in batches and observe the loss function l on the validation set.MSE Whether to decrease to determine if the network based on bilateral residual blocks is overfitting; loss function l MSE The loss function is defined as the mean squared error loss between the output image and the label;

[0013] Step 13, SAR image filtering processing based on loss function;

[0014] Based on the loss function l MSE For single-look complex SAR images X in After speckle filtering, the filtered SAR image X is obtained. f ;

[0015] Step 2: Image segmentation based on edge detection network;

[0016] Step 21: Distinguish between the background region and the target region in the SAR image;

[0017] Using an edge detection network to distinguish input single-look complex SAR images X in Uniform background regions and target structure regions in the image enable single-look complex SAR image X-ray imaging. in Image segmentation;

[0018] The output matrix of the edge detection network is denoted as the edge detection-segmentation image W. seg The W seg Each element in the set takes a value between 0 and 1;

[0019] In training the edge detection network, firstly, using auxiliary information from optical remote sensing images in online maps, and based on the intensity and shadow characteristics in the SAR image, pixels subjectively determined to be target structure regions that need to be retained are labeled as 1, and background pixels are labeled as 0. This is called the binary label image X. 二值 ; For single-view complex SAR images X in and binary labeled image X 二值 Then, data augmentation is performed sequentially using flipping, rotation, and gamma correction to obtain the training dataset.

[0020] Step 22, Training and loss function acquisition based on edge detection network;

[0021] The loss function used during the training phase of the edge detection network is defined as the sum of the six side outputs and the final output compared to the binary label image X. 二值 The weighted cross-entropy, then the training-loss function

[0022] Step 23: Image segmentation based on the training-loss function;

[0023] Image segmentation is performed based on the training loss function L(W,w) to obtain the edge detection-segmented image W. seg ;

[0024] Step 3: SAR image weighted fusion based on linear weighting;

[0025] The speckle suppression gain factor r is the cumulative X f The pixel ratio of the mid-background region to the cumulative X in Mid-background region pixels; specific speckle suppression gain factor

[0026] Then, using the gain factor r and W seg Calculate image weights, filter results for speckle suppression, and X. in Weighted fusion is performed to obtain the fused SAR image X. out The SAR image weighted fusion formula is: Attached Figure Description

[0027] Figure 1 This is a block diagram of a traditional SAR system.

[0028] Figure 2 This is a flowchart of the SAR image speckle suppression method combining network filtering and structure restoration according to the present invention.

[0029] Figure 3 This is a schematic diagram of the DurB-Net-SAR network structure used in the speckle filtering of this invention.

[0030] Figure 4 This is a schematic diagram of the bilateral residual block structure used in this invention.

[0031] Figure 5 This is a schematic diagram showing the selection of different areas on a SAR image of region XX.

[0032] Figure 6 This is a diagram showing the result of processing SAR data for region XX using step one of the methods in this invention.

[0033] Figure 7 This is the final result image of SAR data processing for the XX region using the method of this invention. Detailed Implementation

[0034] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0035] The SAR image speckle suppression method of this invention is performed within a SAR ground system. Specifically, it performs speckle suppression on the single-look complex image output by the Class 1A processor in the SAR signal processor. See [link to related documentation]. Figure 2As shown, the present invention provides a SAR image speckle suppression method combining network filtering and structure restoration, comprising the following steps:

[0036] Step 1: SAR image speckle filtering;

[0037] Step 11, Acquisition of SAR Images;

[0038] The single-look complex image output by the 1A-level processor in the SAR signal processor is denoted as the single-look complex SAR image X. in The single-look complex SAR image X is represented in matrix form. in for:

[0039]

[0040] N a Indicates the number of points in the image orientation;

[0041] N r Indicates the distance to points in the image;

[0042] q is a variable that takes values ​​from 1 to N. a , indicating the direction to the q-th point;

[0043] p is a variable whose value ranges from 1 to N. r , which represents the distance to the p-th point;

[0044] Represents the pixel in the first row and first column of the SAR image;

[0045] Represents the pixel in the 1st row and pth column of the SAR image;

[0046] The Nth row of the SAR image r The number of pixels in a column;

[0047] Represents the pixel in the q-th row and 1-th column of the SAR image;

[0048] Represents the pixel in the q-th row and p-th column of the SAR image;

[0049] Represents the Nth row of the SAR image r The number of pixels in a column;

[0050] Represents the Nth SAR image a The pixel in the first column of the row;

[0051] Represents the Nth SAR imagea The pixel in row p;

[0052] Represents the Nth SAR image a row N r The number of pixels in a column.

[0053] Step 12: Obtain the loss function for the SAR image based on the bilateral residual block;

[0054] In this invention, a DurB-Net-SAR (Dual Residual Block Network for SAR) network structure based on bilateral residual blocks is adopted. The prototype of this network is DurB-Net proposed by Xing Liu et al. in 2019 in "Dual Residual Networks Leveraging the Potential of Paired Operations for Image Restoration," which has been verified to achieve excellent results in optical image restoration tasks. Figure 3 As shown, the input and output parts of DurB-Net-SAR each contain two ordinary convolutional layers. Except for the final output layer, the other three layers all have batch normalization layers added before the ReLU activation function layers. Compared to DurB-Net, DurB-Net-SAR removes the residual connections between the input and output, making it more suitable for non-additive speckle distributions. The core of the network consists of six cascaded bilateral residual blocks (DuRBs). Figure 4 The internal structure of the l-th bilateral residual block is given. Each bilateral residual block contains four layers, where the first two layers are ordinary convolutional layers, and the last two layers are... and For dilated convolutional layers, the first layer of the l-th bilateral residual block. The expansion rate is large, and the second layer of the l-th bilateral residual block The dilation rate is small. The specific dilation rate settings are shown in Table 1. The design based on the bilateral residual block network follows the principle of small convolutional kernels paired with deep networks. There are a total of 28 layers involved in convolutional operations, each layer includes 32 convolutional kernels, the basic size is 3×3, and the convolutional stride is 1.

[0055] Table 1. DurB Expansion Rate Parameter Settings

[0056]

[0057] During training, the SAR image dataset is divided into a SAR image training set and a SAR image validation set. The SAR image training set is denoted as... SAR image validation set denoted as Input the data into the DurB-Net-SAR network in batches and observe the loss function on the validation set. MSE Whether the loss function decreases is used to determine if the DurB-Net-SAR network is overfitting. MSE It is defined as the mean squared error loss between the output image and the label (MSE).

[0058] Specific loss function l MSE for:

[0059]

[0060] m represents the total number of images input into the network in a single iteration.

[0061] i is the identifier of any image in the network input once.

[0062] L W W represents the width of the image, where L is the image width. W Maximum width.

[0063] w is the image width L W The pixels on the surface.

[0064] L H H represents the height of the image, where L is the image height. H The maximum height.

[0065] h is the image height L H The pixels on the surface.

[0066] Y i (w,h) represents the pixel at (w,h) in the i-th input image.

[0067] X i (w,h) represents the pixel at (w,h) in the label of the i-th input image.

[0068] f Θ This represents the current network function, and Θ represents the network parameters.

[0069] Step 13, SAR image filtering processing based on loss function;

[0070] Based on the loss function l MSE For single-look complex SAR images X in After speckle filtering, the filtered SAR image X is obtained. f It can be represented in matrix form as follows:

[0071]

[0072] This represents the pixel in the first row and first column of the speckle filtering result;

[0073] This represents the pixel in the first row and p column of the speckle filtering result;

[0074] The first row, Nth row, represents the results of speckle filtering. r The number of pixels in a column;

[0075] This represents the pixel in the q-th row and 1-th column of the speckle filtering result;

[0076] This represents the pixel in the q-th row and p-th column of the speckle filtering result;

[0077] The Nth row of the coherence speckle filtration result represents the result of the qth row. r The number of pixels in a column;

[0078] The Nth index represents the result of speckle filtration. a The pixel in the first column of the row;

[0079] The Nth index represents the result of speckle filtration. a The pixel in row p;

[0080] The Nth index represents the result of speckle filtration. a row N r The number of pixels in a column.

[0081] In constructing a bilateral residual block network, a SAR simulation dataset is used instead of a commonly used optical simulation dataset to train the speckle suppression network. The trained model can handle the imaging process through echo simulation and range Doppler imaging algorithms. Compared with commonly used optical simulation datasets, the SAR simulation data obtained in this way is closer to the distribution of real test SAR images.

[0082] To address the issue of significant differences in data features between the training set and the test images, a SAR simulation dataset is used instead of the commonly used optical simulation dataset to train the speckle suppression network. The trained model is then used to filter out speckles from the images to be processed.

[0083] The SAR simulation dataset uses SAR images as labels, and the corresponding noisy input is obtained through echo simulation and range-Doppler imaging algorithms. First, each pixel of the label data is considered a scattering unit on the imaging surface, and a random phase shift is added to each pixel; the pixel's amplitude represents the backscattering coefficient of that scattering unit. In the echo simulation, to simulate the real imaging process, each resolution cell contains 4×4 scattering points, so the width and height of the processed image are both one-quarter of the original. Finally, the label data is resampled to ensure that the input size matches the label size.

[0084] Step 2: Image segmentation based on edge detection network;

[0085] Step 21: Distinguish between the background region and the target region in the SAR image;

[0086] In this invention, an edge detection network is used to distinguish the input single-look complex SAR image X. in Uniform background regions and target structure regions in the image enable single-look complex SAR image X-ray imaging. in Image segmentation. The input to the edge detection network is a single-look complex SAR image X. in With the same input as in step one, the output matrix of the edge detection network is denoted as edge detection-segmentation image W. seg The W seg Each element in the set takes a value between 0 and 1.

[0087] In training the edge detection network, firstly, using auxiliary information from optical remote sensing images in online maps, and based on features such as intensity and shadows in the SAR image, pixels in target structure regions that are subjectively determined to be to be retained are marked as 1, and pixels in background regions are marked as 0. This is called the binary label image X. 二值 For single-look complex SAR images X in and binary labeled image X 二值 Then, data augmentation is performed sequentially using flipping, rotation, and gamma correction to obtain the training dataset.

[0088] Step 22, Training and loss function acquisition based on edge detection network;

[0089] In this invention, the DexiNed edge detection network, based on dense connections and upsampling, proposed by Xavier Soria et al. in "Dense Extreme Inception Network: Towards a Robust CNN Model for Edge Detection" in 2020, is used for target and background segmentation of images. Similar to the edge detection problem, the segmentation of target and background regions can be regarded as a pixel binary classification problem with imbalanced sample distribution. DexiNed obtains multi-scale edge features of the image through multiple densely connected encoders (Dexi) and upsampling blocks (UB). The encoder receives the input image, and the main part is a 26-layer cascaded 3×3 convolutional layer, including batch normalization layers and ReLU activation function layers. Due to the depth of the network, in order to prevent the loss of edge detail adjustment, a large number of skip connections containing 1×1 convolutions are used in the network, hence the name densely connected encoder. The cascaded convolutional layers are divided into 6 modules by the top skip connections, where the 2nd, 3rd, and 4th modules end with max pooling layers with a kernel size of 3×3 and a stride of 2. Each of the six main modules extracts a set of feature maps at different scales and feeds them into the side upsampling block to restore the original resolution, resulting in a side sub-output image. The upsampling block consists of two sub-blocks, each containing a 1×1 convolutional layer and an s×s deconvolutional layer, where s is the factor by which the feature map input to the upsampling layer is scaled down compared to the original input image. Finally, the six sub-output images are fused together by a 1×1 convolutional layer into a single image with values ​​in the range [0,1].

[0090] The loss function used during the training phase of the edge detection network is defined as the sum of the six side outputs and the final output compared to the binary label image X. 二值 Given the weighted cross-entropy, the training-loss function L(W,w) is:

[0091]

[0092] W is the image width L W Maximum width.

[0093] w is the image width L W The pixels on the surface.

[0094] δ n The weight of the nth output image is given in Table 2.

[0095] l n (W, w) n ) represents the cross-entropy between the nth side output image and the binary label image.

[0096] w nThis represents the number of pixels in the nth output image.

[0097] β represents the proportion of background pixels, and Y + The set of target pixels, Y - A collection representing the pixels of the background area.

[0098] σ(y j =1|X in ;W,w n This is the accumulation of the six sides of the target pixel.

[0099] y j Let the j-th pixel be the target pixel.

[0100] j is Y + The identifier of any target pixel in the dataset.

[0101] σ(y η =0|X in ;W,w n This is the accumulation of the six sides of the background region pixels.

[0102] y η Let the ηth pixel be a background pixel.

[0103] η is Y - The identifier of any pixel in the background region.

[0104] Table 2 DexiNed Loss Function Parameter Settings

[0105] n <![CDATA[Weight δ n > 1 0.4 2 0.4 3 0.6 4 0.6 5 0.5 6 0.5 7 1.3

[0106] Step 23: Image segmentation based on the training-loss function;

[0107] Image segmentation is performed based on the training loss function L(W,w) to obtain the edge detection-segmented image W. seg It can be represented in matrix form as follows:

[0108]

[0109] Represents the pixel in the first row and first column of the image segmentation result;

[0110] This represents the pixel in the first row and p-th column of the image segmentation result;

[0111] The first row and Nth row of the image segmentation results r The number of pixels in a column;

[0112] Represents the pixel in the q-th row and 1-th column of the image segmentation result;

[0113] Represents the pixel in the q-th row and p-th column of the image segmentation result;

[0114] The Nth row of the image segmentation result represents the q-th row. r The number of pixels in a column;

[0115] The Nth element representing the image segmentation result a The pixel in the first column of the row;

[0116] The Nth element representing the image segmentation result a The pixel in row p;

[0117] The Nth element representing the image segmentation result a row N r The number of pixels in a column.

[0118] Step 3: SAR image weighted fusion based on linear weighting;

[0119] In this invention, the speckle suppression gain factor r is the cumulative X f The pixel ratio of the mid-background region to the cumulative X in The gain factor of the network is obtained by calculating the ratio of the mean values ​​of the background region pixels in the image before and after speckle suppression.

[0120] The specific formula for calculating the speckle suppression gain factor r is as follows:

[0121]

[0122] X in (u,v) represents the single-look complex SAR image X. in The pixel value at (u,v).

[0123] X f (u,v) represents the filtered SAR image X. f The pixel value at (u,v).

[0124] Y - A collection representing the pixels of the background area.

[0125] In this invention, the gain factor r and W are utilized. seg Calculate image weights, filter results for speckle suppression, and X. in Weighted fusion is performed to obtain the fused SAR image X. out The SAR image weighted fusion formula is:

[0126]

[0127] This represents the Hadamard product operator.

[0128] X out This is a SAR image after speckle suppression.

[0129] The main differences between the method of this invention and other deep learning methods are as follows: the method of creating the training set of the speckle suppression network is changed, and image fusion post-processing is added in addition to the general network filtering steps.

[0130] Regarding training data, the method of this invention uses real SAR images instead of optical images as labels, and simulates the real imaging process through echo simulation and range Doppler imaging algorithms to obtain the corresponding input images. Compared with commonly used optical simulation datasets, the SAR simulation data obtained in this way is closer to the distribution of real test images.

[0131] The purpose of adding a post-processing step in image fusion is to restore the target structure damaged during the network filtering process through image fusion. The fusion principle is to make the background area in the fused result more similar to the filtered result, while the target area retains more of its original structure. This method calculates the weighted fusion weights by training a neural network to output a probability matrix of each pixel being in the target area.

[0132] In addition, the method also has the following advantages in the process of speckle suppression of SAR images: (1) It optimizes the composition of the dataset and improves the generalization performance of the model on real SAR images. (2) It improves the structural preservation ability of the speckle suppression network by using image fusion post-processing and reduces the loss of target information. (3) It decomposes the speckle suppression task into two parts: filtering and structural restoration, which reduces the learning difficulty of a single network. (4) The method does not depend on known system parameters and can be adapted to test images with different parameters by replacing the training data.

[0133] Example 1

[0134] This embodiment uses TerraSAR-X satellite data for training and testing. The image resolution is between 0.8 meters and 1.4 meters, and includes both single-view and multi-view images.

[0135] The dataset used in speckle filtering included 4096 pairs of 256×256 image patches as the training set and 128 pairs of 512×512 image patches as the validation set. The network was trained in PyTorch 1.7.1, using default parameter initialization. The Adam algorithm optimization parameters were set to their default values ​​of 0.9 and 0.999, respectively, and the smoothing term was set to 1×10⁻⁶. -8 Each time, 8 samples of 256×256 are read. The initial learning rate is set to 0.001, and then it is reduced to half of the original rate every 20 rounds, for a total of 100 rounds of training.

[0136] The training set used in image segmentation processing contained 800 pairs of original samples of size 256×256, which were augmented to a total of 19,200 pairs. An additional 16 pairs of samples of size 1024×1024 were used as a validation set. The network was trained in PyTorch 1.7.1, using the Adam algorithm to optimize the parameters. The parameters of the 3×3 convolutional kernels were initialized using a Xavier normal distribution, and the parameters of the 1×1 convolutional kernels were initialized using a normal distribution with a mean of 0. The biases were initialized to 0. The learning rate was a constant 0.0001 during training, and 32 training samples were input to the network each time. To prevent overfitting, a weight decay coefficient of 0.0001 was set. The network was trained for a total of 20 epochs.

[0137] The test image was selected from a SAR image slice of the XX region, with a resolution of 1 meter. Figure 5 The image shows the original test SAR image, with six detection regions selected. Table 3 lists the corresponding quantization metrics. Selections B, F, and C calculate the equivalent number of views for the selected regions, measuring the smoothness of the image. EPD calculates the edge preservation index for the entire image; a value closer to 1 indicates stronger edge structure preservation capability. Selections D and E calculate the change in target clutter ratio for the selected regions; a smaller change indicates less impact of the algorithm on strongly scattering targets and better target structure preservation.

[0138] Table 3 Evaluation of SAR data filtering results in the XX region

[0139] algorithm Select B Select F Select C EPD Select D Select E Original image 0.9594 0.9335 0.8197 — — — Filtering results 8.0699 9.8316 5.3058 0.4540 3.2790 2.5214 Final result 8.0350 9.8316 5.2799 0.7468 0.0683 0.2311

[0140] Comparing the data in Table 3, it can be seen that the equivalent number of views in the uniform region before and after image fusion only decreases slightly, while the edge preservation index increases by 60%, and the change in the target clutter ratio also decreases significantly. This objectively proves the effectiveness of the method proposed in this invention and verifies that image segmentation and post-fusion processing can effectively improve the target structure preservation ability of the algorithm while maintaining the network's speckle suppression ability. Through comparison... Figure 5 , Figure 6 , Figure 7 , Figure 7 and Figure 6 There is almost no difference on a relatively uniform sea surface, while Figure 5 Isolated strong scatterers selected in boxes D and E are in Figure 6 The size of the line within the box A is noticeably smaller. Figure 6 The middle part also becomes noticeably thinner, while Figure 7 No similar changes were observed.

Claims

1. A method for suppressing speckle in SAR images by combining network filtering and structure restoration, characterized in that... It includes the following steps: Step 1: SAR image speckle filtering; Step 11, Acquisition of SAR Images; The single-look complex image output by the 1A-level processor in the SAR signal processor is denoted as the single-look complex SAR image. ,and The value of q ranges from 1 to N. a N a This represents the number of points in the image orientation; the value of p ranges from 1 to N. r N r This represents the distance to points in the image; therefore, the image In It refers to the pixel in the q-th row and p-th column of the SAR image; Step 12: Obtain the loss function for the SAR image based on the bilateral residual block; The network structure based on bilateral residual blocks contains two ordinary convolutional layers in both the input and output parts. Except for the final output layer, the other three layers all have batch normalization layers added before the ReLU activation function layer. The core of the network is the six cascaded bilateral residual blocks in the middle. During training, the SAR image dataset is divided into a SAR image training set. and SAR image verification set Input the data into the network based on bilateral residual blocks in batches and observe the loss function on the validation set. Whether to decrease is used to determine if the network based on bilateral residual blocks is overfitting; m represents the total number of images input into the network in a single input; It is the identifier of any image in the network input at a single time; The width of the image. Image width Maximum width; Image width pixels on; The height of the image. Image height Maximum height; Image height pixels on; For the first Zhang input image Pixels at that location; For the first Zhang Input Image The corresponding pixel in the label; Indicates the current network function. Indicates network parameters; Step 13, SAR image filtering processing based on loss function; Based on the loss function For images After speckle filtering, the filtered image is obtained. ,and ; Step 2: Image segmentation based on edge detection network; Step 21: Distinguish between the background region and the target region in the SAR image; Use an edge detection network to distinguish the input image Uniform background region and target structure region in the image to achieve image Image segmentation; The output matrix of the edge detection network is denoted as the edge detection-segmentation image. The Each element in the set takes a value between 0 and 1; Step 22, Training and loss function acquisition based on edge detection network; The loss function used during the training phase of the edge detection network is defined as the sum of the six side outputs and the final output image with the binary label image. The weighted cross-entropy, then the training-loss function is: , No. The cross-entropy between the side output image and the binary label image is ； For the first Weights of the output image; Indicates the first The number of pixels in the output image; This indicates the percentage of background pixels, and , The set representing the target pixels. A collection representing the pixels of the background area; This is the accumulation of the six sides of the target pixel; For the first The pixel is the target pixel; for The identifier of any target pixel in the array; This is the accumulation of pixels on all six sides of the background area; For the first The pixels are background pixels; for The identifier of any pixel in the background region; Step 23: Image segmentation based on the training-loss function; Based on the training-loss function Image segmentation is performed to obtain the edge-detected segmented image. ,and ; Step 3: SAR image weighted fusion based on linear weighting; speckle suppression gain factor It is to accumulate The cumulative pixel ratio in the mid-background area Pixels in the middle background area; Specific speckle suppression gain factor ; Represents single-look complex SAR image middle Pixel value at; Represents the SAR image after filtering. middle Pixel value at; A collection representing the pixels of the background area; Using gain factor and Calculate image weights and the filtering results for speckle suppression. Weighted fusion is performed to obtain the fused SAR image. ,and ; This represents the Hadamard product operator; This is a SAR image after speckle suppression.

2. The SAR image speckle suppression method combining network filtering and structure restoration according to claim 1, characterized in that: In the bilateral residual block of step 12, following the principle of pairing small convolutional kernels with deep networks, there are a total of 28 layers involved in convolution operations. Each layer includes 32 convolutional kernels with a basic size of 3×3 and a convolutional stride of 1.

3. The SAR image speckle suppression method combining network filtering and structure restoration according to claim 1, characterized in that: In the edge detection of step 22, each of the six main modules extracts a set of feature maps at different scales and feeds them into the side upsampling block to restore the original resolution, resulting in a side sub-output image. The upsampling block consists of two sub-blocks, each containing a 1×1 convolutional layer and an s×s deconvolutional layer, where s is the factor by which the feature map of the input upsampling layer is scaled down compared to the original input image. Finally, the six sub-output images are fused into an image with values ​​in the range [0,1] through a 1×1 convolutional layer.

4. The SAR image speckle suppression method combining network filtering and structure restoration according to claim 1, characterized in that: It is to perform speckle suppression on the single-view complex image output by the 1A-level processor in the SAR signal processor.

Images