Collaborative optimization method for speckle noise suppression and structure maintenance of SAR image

Abstract

The invention discloses a collaborative optimization system and method for speckle noise suppression and structure maintenance of an SAR (Synthetic Aperture Radar) image, and the system consists of a structure extraction module based on an edge detection network, a feature fusion module based on convolution and a speckle noise suppression module based on a bilateral residual block network, and is stored in a 1B-level signal processor; the method comprises the following steps: firstly, inputting an original SAR image Xi with speckle noise into image structure information XSi extracted from an Stu module; performing feature fusion on the Xi and the XSi through convolution and an attention mechanism to obtain information UXi after feature fusion; and finally, after UXi is subjected to speckle removal through a Dep module, an SAR image with speckle noise suppression and structure maintenance collaborative optimization is obtained. According to the method, coupling between a speckle noise suppression module and a structure extraction module is utilized, and edge and texture details of an SAR image are better reserved while speckles are removed through guidance of structure information.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and is a collaborative optimization method for suppressing speckle noise and preserving structure of synthetic aperture radar (SAR) images based on deep learning. Background Art

[0002] As an advanced imaging technology, SAR system can obtain high-resolution images of the surface under all-weather and all-day conditions, so it is widely used in military monitoring, environmental monitoring, geological exploration and other fields. However, due to the influence of factors such as Doppler effect, beam interference and surface scattering characteristics during the imaging process, SAR images usually produce a unique noise phenomenon - speckle noise. This noise is caused by the coherence of radar signals, which manifests as high-frequency speckle noise with random distribution in the image, seriously affecting the visual effect of SAR images and subsequent image analysis.

[0003] The first edition of March 2014, Harbin Institute of Technology Press, "Radar Imaging Technology" edited by Liu Yongtan, pages 109, 113-114 pointed out that the synthetic aperture radar (SAR) is installed on a moving platform, and transmits and receives pulses at a certain repetition frequency to form an echo signal. The structural block diagram of the SAR system is shown in the figure below. Figure 1 As shown in the figure, the SAR system includes three parts: the onboard radar system, the satellite platform and the data downlink system, and the ground system. The synthetic aperture radar imaging processing is completed in the ground system. The ground system receives the echo signal sent by the satellite platform and the data downlink system through the ground receiving station. The echo signal is imaged and processed by the SAR signal processor to obtain a SAR image; the SAR image is stored in the backup operating system.

[0004] The presence of speckle noise poses a huge challenge to the processing of SAR images. Although traditional denoising methods, such as mean filtering, median filtering and wavelet transform, can suppress speckle noise to a certain extent, they often lead to the loss of detail information in the image, especially in the edge and texture areas. Since the structural information in SAR images is crucial in tasks such as object recognition and target detection, how to effectively remove speckle noise while retaining the edge and detail information in the image has become a problem to be solved in the field of SAR image processing.

[0005] In recent years, with the rapid development of deep learning technology, researchers have proposed a series of deep learning-based SAR image despeckling methods. These methods use the powerful feature extraction capabilities of neural networks to automatically learn complex patterns in images, thereby better preserving the structural information of the image while removing noise. However, existing methods still face great challenges when dealing with inhomogeneous areas (such as strong scatterers such as buildings and vehicles), which can easily lead to blurred edges or loss of details. Summary of the invention

[0006] Speckle noise is a significant problem in SAR image processing because it destroys texture details, thus limiting the accuracy of its high-precision applications. The impact of speckle noise on SAR images is related to its structural characteristics. Images are usually divided into homogeneous regions, inhomogeneous regions, and extremely inhomogeneous regions. In the latter two regions, structures such as edges and strong point targets will produce more complex speckle distributions than in homogeneous regions. Despeckling filters based on multiplicative models perform well in homogeneous regions, but perform poorly in the other two regions. In order to solve the problem of loss of image structural information when suppressing speckle noise, this paper proposes a SAR image despeckling collaborative optimization method based on structure preservation, which utilizes the coupling between the despeckling module and the structure extraction module. Guided by structural information, the proposed method can better retain edge and texture details while removing speckles.

[0007] A collaborative optimization system for SAR image speckle noise suppression and structure preservation of the present invention, the system is composed of a structure extraction module based on an edge detection network, a feature fusion module based on convolution and a speckle noise suppression module based on a bilateral residual block network;

[0008] The Stu module first extracts image features of different scales through the encoder block to generate multi-scale structure maps; secondly, it improves the feature map resolution through the upsampling block to generate a refined structure map; thirdly, it fuses the intermediate structure maps generated by all upsampling blocks through the structure fusion layer, and generates the final structure extraction feature map through the learned filter;

[0009] The first aspect of the Fus module is to process the SAR image through convolution operation to generate the SAR preliminary feature map; the second aspect is to process the SAR image through convolution operation to generate the attention matrix; the third aspect is to perform weighted fusion of the SAR preliminary feature map and the attention matrix; the fourth aspect is to generate higher quality fusion features through convolution fusion;

[0010] The Dep module first extracts image features preliminarily through an input layer consisting of a convolutional layer, a batch normalization layer, and a RELU activation function; secondly, it suppresses speckle noise through a dual residual block consisting of convolution operations with two different receptive fields, while achieving global feature extraction and local detail preservation; thirdly, it maps the denoising result back to the image through an output layer consisting of a convolutional layer and a hyperbolic tangent activation function, and generates the final speckle noise suppression result.

[0011] The collaborative optimization system for SAR image speckle noise suppression and structure preservation of the present invention is used to prevent the loss of image structure information when suppressing speckle noise. The collaborative optimization comprises the following steps:

[0012] Step 1: Obtain the SAR image to be processed;

[0013] The Stu-Fus-Dep collaborative model obtains any single-view amplitude image, i.e., a SAR image with speckle noise, from a level 1B signal processor, denoted as the image to be processed X i ;

[0014] Step 2: Apply Stu network to extract feature information of digital binary image;

[0015] Step 21, processing the SAR image in the Stu module;

[0016] The Stu network includes an encoder block, an upsampling block, and a structure fusion layer based on the DexiNed architecture;

[0017] The encoder block is used to extract X at different scales i The characteristics of , denoted by AX i ;

[0018] The upsampling block receives AX i , using convolution and transposed convolution to generate an intermediate structure map, denoted as BX i ;

[0019] The structure fusion layer maps the intermediate structure generated by all upsampling blocks to BX i Fusion is performed to generate the final structure extraction feature map, denoted as CX i ;

[0020] Step 22, obtaining feature information of the digitized binary image;

[0021] According to the trained Stu model F Stu About CX i Processing is performed to obtain a binary structure label image XS i ;

[0022] Step 3: Apply the Fus module to complete feature fusion based on convolution and attention;

[0023] Step 31, feature extraction;

[0024] X i Through the convolution operation of the 3×3 convolution layer, the preliminary feature map X of the image is generated. i ′;

[0025] Step 32, generate an attention matrix;

[0026] XS i Through a 1×1 convolution operation, the spatial attention matrix XS is generated i ′;

[0027] Step 33, feature weighting;

[0028] X i ′ and XS i 'Multiply element by element to form a weighted feature map XXS i ′;

[0029] Step 34, second weighted fusion;

[0030] XXS i ′Fuse again through the convolutional layer to obtain the output feature information after weighted fusion, recorded as UX i ;

[0031] Step 4: Speckle noise suppression is completed through Dep;

[0032] Step 41, processing the feature fusion information in the Dep module;

[0033] The input layer initially extracts UX i The characteristics of i ′;

[0034] Use the first connected 6 double residual blocks to UX i ' is processed to obtain the denoising result RX i ;

[0035] The output layer will be RX i Mapped back to the image space, denoted as OX i ;

[0036] Step 42, obtaining a speckle noise suppression result;

[0037] According to F Dep OX i After processing, the speckle noise suppression result X is obtained. o ;

[0038] The F Dep It comes from the output of the trained Dep module.

[0039] Technical effects of the present invention:

[0040] ① Efficient removal of speckle noise: The present invention utilizes a bilateral residual block network to suppress speckle noise, which significantly reduces noise interference, improves the clarity and readability of SAR images, and makes subsequent image analysis more accurate and reliable.

[0041] ② Structural information preservation and enhancement: By introducing a structure extraction network, the present invention can retain the edges and details in the image while removing noise. This structure preservation function is particularly important for tasks such as ground object recognition and target detection in SAR images, and effectively improves the practicality of the image.

[0042] ③ Reduce edge blur: The collaborative optimization framework effectively avoids the edge blur phenomenon common in traditional denoising methods by combining structure extraction and denoising modules. Experiments show that the edge areas in the denoised image are still clear, ensuring the resolution of the image in strong scatterers and detail areas.

[0043] ④ Construction of a highly adaptable dataset: Using echo simulation and SAR imaging processing technology, a training dataset based on SAR images was constructed, enabling the algorithm to cope with complex ground scattering environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is the block diagram of the SAR platform system structure.

[0045] Figure 2 It is a flow chart of the collaborative optimization method for SAR image speckle noise suppression and structure preservation proposed by the present invention.

[0046] Figure 3 It is the loss function design of the collaborative optimization network for SAR image speckle noise suppression and structure preservation proposed by the present invention.

[0047] Figure 4 This is an example of a SAR image from the Stu network dataset.

[0048] Figure 5 is from Figure 4 The structural information extracted from the Stu network dataset is an example of a labeled image.

[0049] Figure 6 This is an example of a labeled image from the Dep network dataset.

[0050] Figure 7 is Figure 6 The simulated image generated by the echo simulation and imaging processing technology of the SAR system is an example of the speckle noise suppression network dataset image.

[0051] Figure 8 This is the SAR image effect before speckle noise suppression.

[0052] Fig. 9 yes Figure 8 SAR image results after speckle noise suppression. DETAILED DESCRIPTION

[0053] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. The various parameter demonstrations listed are only preferred embodiments of the present invention. It should be noted that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention.

[0054] The present invention relates to a collaborative optimization method for SAR image speckle noise suppression and structure preservation, and the method aims to solve the problem of image structure information loss caused by suppressing speckle noise in the prior art. The collaborative optimization of SAR image speckle noise suppression and structure preservation of the present invention consists of a structure extraction module based on an edge detection network, a feature fusion module based on convolution, and a speckle noise suppression module based on a bilateral residual block network, referred to as the Stu-Fus-Dep collaborative model, and the Stu-Fus-Dep collaborative model is stored in a Class 1B signal processor. The method combines the Dep model and the Stu module in deep learning based on the collaborative optimization of SAR image speckle noise suppression, and through the collaborative optimization of the two, while removing speckle noise, effectively retains the edge and texture details in the SAR image.

[0055] The speckle noise suppression module (i.e., Dep module) based on the bilateral residual block network refers to X.Liu, M.Suganuma, Z.Sun, and T.Okatani, “Dual Residual Networks Leveraging the Potential of Paired Operations for Image Restoration,” in Proc.IEEE CVPR, 2019, pp.7000-7009.

[0056] The structure extraction module based on the edge detection network (i.e., Stu module) refers to X.Soria, E.Riba, and A.Sappa, "Dense Extreme Inception Network: Towards a Robust CNN Model for EdgeD etection," in Proc.IEEE WACV, Snowmass, CO, USA, 2020, pp.1912-1921. It is referred to as the DexiNed architecture.

[0057] Structure extraction module based on edge detection network (Stu module)

[0058] In the present invention, the Stu module firstly extracts image features of different scales through the encoder block to generate a multi-scale structure map; secondly, it improves the feature map resolution through the upsampling block to generate a refined structure map; thirdly, it fuses the intermediate structure maps generated by all upsampling blocks through the structure fusion layer, and generates the final structure extraction feature map through the learned filter.

[0059] Convolution-based feature fusion module (Fus module)

[0060] In the present invention, the Fus module firstly processes the SAR image through a convolution operation to generate a SAR preliminary feature map; secondly, processes the SAR image through a convolution operation to generate an attention matrix; thirdly, the SAR preliminary feature map and the attention matrix are weightedly fused; fourthly, higher quality fusion features are generated through convolution fusion.

[0061] Speckle noise suppression module (Dep module) based on bilateral residual block network

[0062] In the present invention, the Dep module firstly extracts image features through an input layer composed of a convolution layer, a batch normalization layer and a RELU activation function; secondly, it suppresses speckle noise through a dual residual block composed of convolution operations with two different receptive fields, while achieving global feature extraction and local detail preservation; thirdly, it maps the denoising result back to the image through an output layer composed of a convolution layer and a hyperbolic tangent activation function, and generates the final speckle noise suppression result.

[0063] In the present invention, the SAR image with speckle noise is recorded as the original SAR image.

[0064] The processing flow of the present invention is as follows Figure 2 First, the original SAR image X with speckle noise is i Input into the structure extraction module based on edge detection network, and the extracted image structure information is recorded as XS i ; Then, the image to be processed X i and XS i The feature fusion is performed through convolution and attention mechanism, and the information after feature fusion is recorded as UX i Finally, UX i After despeckle by the speckle noise suppression module based on the bilateral residual block network, the obtained SAR image with coordinated optimization of speckle noise suppression and structure preservation is denoted as X o .

[0065] Construction of structure extraction module based on edge detection network, namely Stu model

[0066] Construction step A, obtain training images;

[0067] A number of SAR single-view complex images with spots are selected from the 1B signal processor as training images, denoted as the training image set TRIM = {image1, image2, …, image a ,…,image A}.

[0068] Image1 represents the first training image.

[0069] image2 represents the second training image.

[0070] image a Represents the a-th training image.

[0071] image A Represents the last training image.

[0072] For the convenience of explanation, image a It is also called any training image. a The pixel matrix is ​​expressed as

[0073] N a is the total number of pixels in azimuth.

[0074] N r is the total number of pixels in the distance direction.

[0075] χ=1,2,3,…,N a , represents the xth pixel in the azimuth direction.

[0076] γ=1,2,3,…,N r , represents the γth pixel point in the distance direction.

[0077] trx χ,γ Represents the pixel at the xth row and yth column of the training image.

[0078] trx 1,γ Indicates the Nth row of the training image r Column of pixels.

[0079] trx χ,1 Represents the Nth training image a The pixel at row and column 1.

[0080] Represents the Nth training image aLine No. N r The pixel in the column is also the last pixel in the training image.

[0081] Construction step B, distinction between structure area and background area;

[0082] With the help of the auxiliary information of optical remote sensing images in the online map, according to the characteristics of SAR images, TRIM={image1,image2,…,image a ,…,image A The pixels in the structure area to be retained are marked as 1, and the pixels in the background area are marked as 0, and the training binary label image set is obtained, which is denoted as TRI, and TRI = {ri1, ri2, …, ri a ,…,ri A}.

[0083] ri1 represents the training binary label image after the training image image1 is distinguished between the structure area and the background area.

[0084] ri2 represents the training binary label image after the training image image2 is distinguished between the structure area and the background area.

[0085] ri a Represents the training image image a Training binary label image after distinguishing structure area and background area.

[0086] ri A Represents the training image image A Training binary label image after distinguishing structure area and background area.

[0087] Construction step C, build the Stu training network based on the DexiNed architecture;

[0088] In the present invention, the DexiNed architecture includes multiple sets of filters, which receive an image as input and then predict a structure map with the same resolution.

[0089] In the present invention, an encoder block, an upsampling block and a structure fusion layer are provided in the Stu training network according to the DexiNed architecture.

[0090] The encoder block contains an encoder consisting of 6 convolutional blocks to extract image features of different scales and generate multi-scale structure maps. Each encoder consists of two convolutional layers with 3×3 convolution kernels, and is equipped with batch normalization and ReLU activation function.

[0091] The upsampling block receives the feature map from the encoder block and generates an intermediate structure map using convolution and transposed convolution.

[0092] The structure fusion layer fuses the intermediate structure maps generated by all upsampling blocks and generates the final structure extraction feature map.

[0093] Construct step D and set the loss function of Stu training network;

[0094] Set the Stu network loss function during training

[0095] In the present invention, the Stu network loss function adopts the BCDN loss function of the network output and the structure label, and the relationship between the constructed Stu module loss function is:

[0096] Represents the loss function required by Stu to train the network.

[0097] Represents the BCDN loss output by the Stu training network.

[0098] Represents the output digitized binary image feature information.

[0099] In the present invention, according to the loss function in the Stu training network TRIM={image1,image2,…,image a ,…,image A} and TRI={ri1,ri2,…,ri a ,…,ri A}After model training, the trained Stu module F is obtained Stu .

[0100] Construction of the speckle noise suppression module of the bilateral residual block network, namely the Dep model

[0101] Construction step 1, obtain training images;

[0102] A number of SAR single-view complex images with spots are selected from the 1B signal processor as training images, denoted as the training image set TRIM = {image1, image2, …, image a ,…,image A}.

[0103] The image a The pixel matrix is ​​expressed as

[0104]

[0105] Construction step 2, obtain the noise-free label dataset;

[0106] TRIM={image1,image2,…,image a ,…,image A} as a noise-free reference tag, and then the echo simulation and imaging processing technology of the SAR system is used to obtain the imaging result of the simulated data through the range-Doppler algorithm, which is recorded as SM_TRIM = {sm1, sm2, …, sm a ,…,sm A}.

[0107] sm1 represents the imaging result of the training image image1 after echo simulation and imaging processing by the SAR system.

[0108] sm2 represents the imaging result of the training image image2 after echo simulation and imaging processing by the SAR system.

[0109] sm a Represents the training image image a The imaging results after echo simulation and imaging processing by the SAR system.

[0110] sm A Represents the training image image A The imaging results after echo simulation and imaging processing by the SAR system.

[0111] Specifically, image a The amplitude is used as the scattering coefficient in the echo simulation, and the imaging result of the simulated data is obtained through the range-Doppler algorithm to generate image a The corresponding simulation data sm a .

[0112] In order to accurately capture the complexity and diversity of ground work scenarios, sm a The echo simulation process uses an image with a size of 1024×1024 a The amplitude of is used as the scattering coefficient in the echo simulation. Considering the complexity of the ground working scene, 4×4 scattering points are placed in each resolution unit to serve the working conditions during the simulation. The simulated image sm after imaging a The size is 256 × 256. Finally, in order to ensure the consistency of the noisy input of the Stu network with the noise-free reference label, the image with a size of 1024 × 1024 is a Downsample to 256×256.

[0113] Construction step 3, build Dep training network based on DuRB architecture;

[0114] In the present invention, the Dep training network based on the DuRB architecture uses dual residual modules and combines large receptive field and small receptive field convolution operations to achieve efficient extraction of global features and local details. Through outer layer residual connections and modular stacking, the Dep training network can retain image texture and details while removing noise.

[0115] In the present invention, the Dep training network consists of an input layer, 6 dual residual blocks and an output layer.

[0116] The input layer consists of a convolutional layer, a batch normalization layer, and a RELU activation function, which is used to preliminarily extract image features.

[0117] The dual residual block uses two convolution operations with different receptive fields. The first convolution uses a large-scale convolution kernel to capture a wider range of information. The second convolution uses a small-sized convolution kernel to preserve fine details in the image. This paired operation design ensures that the network can balance global feature extraction and local detail preservation at the same time.

[0118] The output layer contains a convolutional layer and a hyperbolic tangent activation function to map the denoising result back to the image space.

[0119] Build step 4 and set the loss function of the Dep training network;

[0120] Set the Dep network loss function L during training 训练 .

[0121] In the present invention, the Dep network loss function adopts the mean square error, so the constructed Dep model loss function relationship is

[0122] Indicates that the pixel carries information loss. Calculated sm a Output after Dep network With the noise-free reference label image a The mean square error between

[0123] Indicates the loss of pixel structure information. Quantified Output after passing through Stu network and the noise-free reference label image a Output after passing through Stu network The mean square error between

[0124] λ is and The linear combination coefficient of . In the present invention, λ is taken as 0.5.

[0125] At this point, the Dep model training is completed.

[0126] In the present invention, see Figure 3 As shown in the Dep training network, based on the loss function L 训练 For SM_TRIM = {sm1, sm2, ..., sm a ,…,sm A} and TRIM = {image1,image2,…,image a ,…,image A}After model training, the trained Dep module F is obtained Dep .

[0127] See also Figure 2 As shown, a collaborative optimization method for SAR image speckle noise suppression and structure preservation of the present invention is denoted as Stu-Fus-Dep collaborative model, and the specific steps are as follows:

[0128] Step 1: Obtain the SAR image to be processed;

[0129] In the present invention, the Stu-Fus-Dep collaborative model obtains any single-view amplitude image, i.e., a SAR image with speckle noise, from a level 1B signal processor, and records it as the image to be processed X. i . The X i The available matrix is ​​represented as

[0130] N a is the total number of pixels in azimuth.

[0131] N r is the total number of pixels in the distance direction.

[0132] q=1,2,3,…,N a , represents the qth pixel in the azimuth direction.

[0133] p=1,2,3,…,N r , represents the pth pixel in the distance direction.

[0134] x q,p Represents the pixel at the qth row and pth column of the single-view amplitude image.

[0135] Represents the Nth row of the single-view amplitude image r Column of pixels.

[0136] Represents the Nth a The pixel at row and column 1.

[0137] Represents the Nth a Line No. N r The pixel point of the column is also the last pixel point of the single-view amplitude image.

[0138] Step 2: Apply Stu network to extract feature information of digital binary image;

[0139] Step 21, processing the SAR image in the Stu module;

[0140] In the present invention, the Stu network includes an encoder block, an upsampling block and a structure fusion layer according to the DexiNed architecture.

[0141] The encoder block is used to extract X at different scales i The characteristics of , denoted by AX i .

[0142] The upsampling block receives AX i , using convolution and transposed convolution to generate an intermediate structure map, denoted as BX i .

[0143] The structure fusion layer maps the intermediate structure generated by all upsampling blocks to BX i Fusion is performed to generate the final structure extraction feature map, denoted as CX i .

[0144] Step 22, obtaining feature information of the digitized binary image;

[0145] According to the trained Stu model F Stu About CX i Processing is performed to obtain a binary structure label image XS i .

[0146] Step 3: Apply the Fus module to complete feature fusion based on convolution and attention;

[0147] Step 31, feature extraction;

[0148] X i Through the convolution operation of the 3×3 convolution layer, the preliminary feature map X of the image is generated. i ′. This operation can not only effectively capture the local spatial information of the image, but also enhance the expressiveness of the features while keeping the feature size of the input and output consistent.

[0149] Step 32, generate an attention matrix;

[0150] XS i Through a 1×1 convolution operation, the spatial attention matrix XS is generated i′. This operation can reduce the feature dimension and realize the spatial weight distribution of features.

[0151] Step 33, feature weighting;

[0152] X i ′ and XS i 'Multiply element by element to form a weighted feature map XXS i ′. This weighted operation can adaptively enhance the feature expression of key areas in the image while weakening the influence of unimportant areas.

[0153] Step 34, second weighted fusion;

[0154] XXS i ′Fuse again through the convolutional layer to obtain the output feature information after weighted fusion, recorded as UX i This operation enhances the ability to capture global information by increasing the receptive field and re-aggregating features, while generating higher quality fusion features.

[0155] Step 4: Speckle noise suppression is completed through Dep;

[0156] Step 41, processing the feature fusion information in the Dep module;

[0157] The input layer initially extracts UX i The characteristics of i ′.

[0158] Use the first connected 6 double residual blocks to UX i ' is processed to obtain the denoising result RX i This operation is used to balance global feature extraction and local detail preservation.

[0159] The output layer will be RX i Mapped back to the image space, denoted as OX i .

[0160] Step 42, obtaining a speckle noise suppression result;

[0161] According to F Dep OX i After processing, the speckle noise suppression result X is obtained. o .

[0162] The F Dep It comes from the output of the trained Dep module.

[0163] The main contributions and innovations of the method proposed in this invention include three parts: network structure, network loss function and data set and construction.

[0164] In terms of network structure, the present invention designs a collaboratively optimized network structure that combines a speckle noise suppression module and a structure extraction module. Under the guidance of structural information, the proposed method can better preserve edge and texture details while suppressing speckle noise.

[0165] In terms of loss function, the present invention defines a new loss function that takes into account both spatial details and structural information.

[0166] In terms of data set construction, the present invention proposes a new method for constructing SAR data sets, which utilizes echo simulation and SAR system imaging processing technology. This method uses the amplitude of real SAR images as the scattering coefficient in echo simulation, obtains the imaging results of simulated data through the range Doppler algorithm, and thus generates a data set based on real SAR images.

[0167] Example 1

[0168] This embodiment uses TerraSAR-X satellite data for training and testing. The resolution of the image is between 0.8 meters and 1.4 meters, including both single-view images and multi-view images.

[0169] When the Stu network is trained, 800 pairs of samples with a size of 256×256 are included. One of the pairs is the training set sample of the Stu network. Figure 4 and Figure 5 As shown, Figure 4 is a SAR image (the scale of the image is pixel coordinates), Figure 5 for Figure 4 During the training process, the parameters of the 3×3 convolution kernel are initialized using the Xavier normal distribution, the parameters of the 1×1 convolution kernel are initialized using the normal distribution with a mean of 0, and the bias is initialized to 0. During the training process, the learning rate is set to a constant value of 0.0001, and the number of image samples input to the network each time is 32. In order to prevent overfitting, the weight decay coefficient is set to 0.0001. The network is trained for a total of 20 rounds.

[0170] The dataset used in the Dep network uses 4096 pairs of image slices with an image size of 256×256. One pair is the training set sample of the Dep network. Figure 6 and Figure 7 shown. Figure 6 is the original SAR image, Figure 7 Based on Figure 6 The generated echo simulation and SAR system simulation results. During the training process, the Adam optimization algorithm was used to optimize the parameters. The β1 and β2 parameters of the optimizer were set to 0.9 and 0.999 respectively, and the smoothing term ε was set to 1×10 -8The training process uses an exponential step decay learning rate setting, with an initial value of 0.001, which decays by half every 10 training cycles. When initializing the Dep network parameters, the weight parameters of the convolution kernel are randomly initialized using a Gaussian distribution with a mean of 0 and a standard deviation of 0.02, and the bias parameters are initialized to 0. During the training process, the number of image samples input to the network at one time is 4, and a total of 100 Dep networks are trained.

[0171] The test image is a slice of TerraSAR-X satellite data with a resolution of 1m. Figure 8 is the original SAR image. First, the original image is applied to the Stu module to extract the digital binary image feature information, then the original image and its corresponding digital binary image feature information are put into the Fus module for feature fusion, and then the result of feature fusion is put into the Dep module for speckle noise suppression, and the final speckle noise suppression result can be output. Fig. 9 The test results of the SAR image speckle noise suppression method proposed in the present invention are shown in Table 1. By comparing the two images, it can be found that the method proposed in the present invention can greatly suppress speckle noise, and the image details, especially in the high-feature structure area, are effectively preserved. Table 1 lists the quantitative indicators of the results. ENL stands for equivalent visual number, which is an extension of SAR resolution and indicates the size of the equivalent ground area represented by a pixel. TCR stands for target clutter ratio, and the change in target clutter ratio ΔTCR can reflect the algorithm's structural preservation ability. EPI stands for edge preservation index, which is an indicator used to evaluate the edge preservation ability of image processing algorithms.

[0172] Table 1 Quantitative evaluation results of SAR image speckle noise suppression effect

[0173]

[0174] According to the results in Table 1, after the speckle noise is suppressed, the equivalent view number of the image is significantly improved, while the target clutter ratio changes less and the edge preservation index is larger, indicating that the algorithm can well suppress speckle noise while retaining the structural information of the image.

Claims

1. A collaborative optimization system for SAR image speckle noise suppression and structure preservation, characterized by: The system consists of a structure extraction module based on edge detection network, a feature fusion module based on convolution, and a speckle noise suppression module based on bilateral residual block network; The Stu module first extracts image features of different scales through the encoder block to generate multi-scale structure mapping; Secondly, the upsampling block is used to improve the feature map resolution and generate a refined structure map. Thirdly, the structure fusion layer is used to fuse the intermediate structure maps generated by all upsampling blocks, and the learned filter is used to generate the final structure extraction feature map. The first aspect of the Fus module is to process the SAR image through convolution operation to generate the SAR preliminary feature map; the second aspect is to process the SAR image through convolution operation to generate the attention matrix; the third aspect is to perform weighted fusion of the SAR preliminary feature map and the attention matrix; the fourth aspect is to generate higher quality fusion features through convolution fusion; The first aspect of the Dep module is to preliminarily extract image features through an input layer consisting of a convolutional layer, a batch normalization layer, and a RELU activation function; Secondly, speckle noise suppression is performed through a dual residual block composed of convolution operations with two different receptive fields, while achieving global feature extraction and local detail preservation; thirdly, the denoising result is mapped back to the image through an output layer composed of a convolution layer and a hyperbolic tangent activation function, and the final speckle noise suppression result is generated.

2. The system for collaborative optimization of SAR image speckle noise suppression and structure preservation according to claim 1, characterized in that: The system is stored in a Class 1B signal processor.

3. The system for collaborative optimization of SAR image speckle noise suppression and structure preservation according to claim 1, characterized in that: The system is used to solve the problem in the prior art that image structure information is lost when suppressing speckle noise.

4. The method of constructing a structure extraction module based on edge detection network has the following steps: Construction step A, obtain training images; A number of SAR single-view complex images with spots are selected from the 1B signal processor as training images, denoted as the training image set TRIM = {image1, image2, …, image a ,…,image A }; Image1 represents the first training image; Image2 represents the second training image; image a represents the a-th training image; image A represents the last training image; The image a The pixel matrix is ​​expressed as N a is the total number of pixels in azimuth; N r is the total number of pixels in the distance direction; χ=1,2,3,...,N a , represents the xth pixel point in the azimuth direction; γ=1,2,3,...,N r , represents the γth pixel point in the distance direction; trx χ,γ represents the pixel at the xth row and the γth column of the training image; trx 1,γ Indicates the Nth row of the training image r Column pixels; trx χ,1 Represents the Nth training image a The pixel in the first column of the row; Represents the Nth training image a Line No. N r The pixel point of the column is also the last pixel point of the training image; Construction step B, distinction between structure area and background area; TRIM={image1,image2,…,image a ,…,image A The pixels in the structure area to be retained are marked as 1, and the pixels in the background area are marked as 0, and the training binary label image set is obtained, which is denoted as TRI, and TRI = {ri1, ri2, …, ri a ,…,ri A }; ri1 represents the training binary label image after the training image image1 is distinguished by the structure area and the background area; ri2 represents the training binary label image after the training image image2 is distinguished by the structure area and the background area; ri a Represents the training image image a The training binary label image after distinguishing the structure area and the background area; ri A Represents the training image image A The training binary label image after distinguishing the structure area and the background area; Construction step C, build the Stu training network based on the DexiNed architecture; In the Stu training network, encoder blocks, upsampling blocks, and structure fusion layers are set according to the DexiNed architecture; The encoder block contains an encoder composed of 6 convolution blocks, which are used to extract image features of different scales and generate multi-scale structure maps; each encoder consists of two convolutional layers with 3×3 convolution kernels, and is equipped with batch normalization and ReLU activation function; The upsampling block receives the feature map from the encoder and generates an intermediate structure map using convolution and transposed convolution; The structure fusion layer fuses the intermediate structure maps generated by all upsampling blocks and generates the final structure extraction feature map; Construct step D and set the loss function of Stu training network; Set the Stu network loss function during training The constructed Stu module loss function relationship is Represents the loss function required by Stu to train the network; represents the BCDN loss output by the Stu training network; Represents the output digital binary image feature information; In Stu training network, based on loss function TRIM={image1,image2,…,image a ,…,image A } and TRI={ri1,ri2,…,ri a ,…,ri A }After model training, the trained Stu module F is obtained Stu .

5. The method for constructing a speckle noise suppression module of a bilateral residual block network has the following steps: Construction step 1, obtain training images; A number of SAR single-view complex images with spots are selected from the 1B signal processor as training images, denoted as the training image set TRIM = {image1, image2, …, image a ,…,image A }; The image a The pixel matrix is ​​expressed as Construction step 2, obtain the noise-free label dataset; TRIM={image1,image2,…,image a ,…,image A } as a noise-free reference tag, and then the echo simulation and imaging processing technology of the SAR system is used to obtain the imaging result of the simulated data through the range-Doppler algorithm, which is recorded as SM_TRIM = {sm1, sm2, …, sm a ,…,sm A }; sm1 represents the imaging result of the training image image1 after echo simulation and imaging processing by the SAR system; sm2 represents the imaging result of the training image image2 after echo simulation and imaging processing by the SAR system; sm a Represents the training image image a Imaging results after echo simulation and imaging processing by the SAR system; sm A Represents the training image image A Imaging results after echo simulation and imaging processing by the SAR system; The image a The amplitude is used as the scattering coefficient in the echo simulation, and the imaging result of the simulated data is obtained through the range-Doppler algorithm to generate image a The corresponding simulation data sm a ; In order to accurately capture the complexity and diversity of ground work scenarios, sm a The echo simulation process uses an image with a size of 1024×1024 a The amplitude is used as the scattering coefficient in the echo simulation; considering the complexity of the ground working scene, 4×4 scattering points are placed in each resolution unit during the simulation to serve the working conditions; the simulated image sm after imaging a The size is 256×256; finally, in order to ensure the consistency of the noisy input of the Stu network with the noise-free reference label, the image with a size of 1024×1024 a Downsample to 256×256; Construction step 3, build Dep training network based on DuRB architecture; The Dep training network consists of an input layer, 6 dual residual blocks, and an output layer; The input layer consists of a convolutional layer, a batch normalization layer, and a RELU activation function, which is used to preliminarily extract image features; The dual residual block uses two convolution operations with different receptive fields. The first convolution uses a large-scale convolution kernel to capture a wider range of information. The second convolution uses a small-sized convolution kernel to preserve fine details in the image. This paired operation design ensures that the network can balance global feature extraction and local detail preservation at the same time. The output layer contains a convolution layer and a hyperbolic tangent activation function to map the denoising result back to the image space; Build step 4 and set the loss function of the Dep training network; Set the Dep network loss function L during training 训练 ; The loss function of the Dep network uses the mean square error, so the loss function relationship of the constructed Dep model is: Indicates that the pixel carries information loss; Calculated sm a Output after Dep network With the noise-free reference label image a The mean square error between Indicates the loss of pixel structure information; Quantified Output after passing through Stu network and the noise-free reference label image a Output after passing through Stu network The mean square error between λ is and The linear combination coefficient of; In the present invention, λ is taken as 0.5; At this point, the Dep model training is completed; In the Dep training network, the loss function L is used to train SM_TRIM={sm1,sm2,…,sm a ,…,sm A } and TRIM = {image1,image2,…,image a ,…,image A }After model training, the trained Dep module F is obtained Dep .

6. A collaborative optimization method for SAR image speckle noise suppression and structure preservation, characterized in that There are the following steps: Step 1: Obtain the SAR image to be processed; The Stu-Fus-Dep collaborative model obtains any single-view amplitude image, i.e., a SAR image with speckle noise, from a level 1B signal processor, denoted as the image to be processed X i ; said X i The available matrix is ​​represented as N a is the total number of pixels in azimuth; N r is the total number of pixels in the distance direction; q=1,2,3,…,N a , represents the qth pixel in the azimuth direction; p=1,2,3,…,N r , represents the pth pixel point in the distance direction; x q,p represents the pixel at the qth row and pth column of the single-view amplitude image; Represents the Nth row of the single-view amplitude image r Column pixels; Represents the Nth a The pixel in the first column of the row; Represents the Nth a Line No. N r The pixel point of the column is also the last pixel point of the single-view amplitude image; Step 2: Apply Stu network to extract feature information of digital binary image; Step 21, processing the SAR image in the Stu module; The Stu network includes an encoder block, an upsampling block, and a structure fusion layer based on the DexiNed architecture; The encoder block is used to extract X at different scales i The characteristics of i ; The upsampling block receives AX i , using convolution and transposed convolution to generate an intermediate structure map, denoted as BX i ; The structure fusion layer maps the intermediate structure generated by all upsampling blocks to BX i Fusion is performed to generate the final structure extraction feature map, denoted as CX i ; Step 22, obtaining feature information of the digitized binary image; According to the trained Stu model F Stu About CX i Processing is performed to obtain a binary structure label image XS i ; Step 3: Apply the Fus module to complete feature fusion based on convolution and attention; Step 31, feature extraction; X i Through the convolution operation of the 3×3 convolution layer, the preliminary feature map X of the image is generated. i ′; Step 32, generate an attention matrix; XS i Through a 1×1 convolution operation, the spatial attention matrix XS is generated i ′; Step 33, feature weighting; X i ' and XS i 'Multiply element by element to form a weighted feature map XXS i ′; Step 34, second weighted fusion; XXS i ′Fuse again through the convolutional layer to obtain the output feature information after weighted fusion, recorded as UX i ; Step 4: Speckle noise suppression is completed through Dep; Step 41, processing the feature fusion information in the Dep module; The input layer initially extracts UX i The characteristics of i ′; Use the first connected 6 double residual blocks to UX i ' is processed to obtain the denoising result RX i ; The output layer will be RX i Mapped back to the image space, denoted as OX i ; Step 42, obtaining a speckle noise suppression result; According to F Dep OX i After processing, the speckle noise suppression result X is obtained. o ; The F Dep It comes from the output of the trained Dep module.

Images