An interpretable method and system for synthetic aperture radar image filtering

Abstract

This invention discloses an interpretable method and system for synthetic aperture radar (SAR) image filtering. The method includes: acquiring SAR images and performing initial segmentation; performing image filtering based on the initial segmentation structure; and performing posterior interpretation on the filtered image to enhance image understanding. This invention constructs an interpretable SAR image filtering method by combining prior guidance with deep learning. It addresses the difficulty of balancing traditional denoising and detail preservation by providing robust denoising for regions of different properties while maintaining detail preservation. It also addresses the "black box" problem of traditional denoising processes by integrating prior-guided mathematical interpretability into the convolutional operations of deep learning, constructing interpretable convolutional filtering. This invention improves the robustness and accuracy of downstream tasks through interpretable SAR image denoising.

Description

Technical Field

[0001] This invention relates to the field of image filtering, and more specifically to an interpretable method and system for synthetic aperture radar image filtering. Background Technology

[0002] Effective noise suppression is crucial for the subsequent interpretation of SAR images. Traditional SAR image processing techniques often neglect the coherence of noise, leading to the loss of important details during filtering. With the advancement of deep learning, image processing techniques have also made significant progress. However, existing deep learning methods do not fully utilize the imaging mechanism of SAR, resulting in a lack of specificity and interpretability in the filtering process. To balance noise reduction and detail preservation, and to address the "black box" problem in filtering, this paper proposes a solution.

[0003] Therefore, in order to improve the robustness of SAR image filtering and the interpretability of the filtering process, new products need to address the contradiction between noise suppression and detail preservation, while ensuring the process is interpretable.

[0004] Currently, to address the key challenge of balancing speckle noise mitigation and image quality preservation in SAR image filtering, prior-guided deep learning can be used to achieve more robust filtering of uniform and non-uniform regions in SAR images. By performing correlation segmentation filtering on the texture and edge contours of different regions, texture blurring and over-smoothing problems can be alleviated, thereby improving image fidelity.

[0005] On the other hand, from the perspective of interpretability: the combination of the global nature of traditional prior-guided filtering and the local nature of deep learning effectively overcomes the limitations of traditional filters that rely too much on local spatial correlations, and has a certain effect on enhancing image details. The filtering process can be explained and visualized through deep learning. Summary of the Invention

[0006] To address the aforementioned technical challenges, this invention proposes a robustness and interpretability optimization method for SAR image filtering. It aims to resolve the conflict between noise suppression and detail preservation in existing methods by combining prior guidance and deep learning techniques. This method uses prior guidance to differentiate between uniform and non-uniform regions in SAR images. Combined with the local feature extraction capabilities of deep learning, it performs correlation segmentation filtering on textures and edge contours in different regions, effectively alleviating texture blurring and over-smoothing, and improving image fidelity. Simultaneously, deep learning visualization techniques enhance the interpretability of the filtering process, overcoming the limitations of traditional filters that overly rely on local spatial correlations. This achieves a balance between noise reduction and detail preservation, providing a higher-quality image foundation for subsequent SAR image interpretation tasks.

[0007] To achieve the above objectives, the present invention provides an interpretability method for synthetic aperture radar image filtering, comprising the following steps:

[0008] Acquire synthetic aperture radar images and perform initial segmentation;

[0009] Based on the preliminary structural division, image filtering is performed.

[0010] A posteriori interpretation is performed on the filtered image to enhance the understanding of the image.

[0011] Preferably, the initial segmentation step includes: based on the prior guidance of traditional filtering algorithms, combining the mean-shift algorithm with Markov random fields to construct a mean-shift-Markov random field filter in the joint space-spectral domain; the MS-MRF filter performs region segmentation of the image through an upward search strategy of correlation density peaks to achieve denoising and detail preservation.

[0012] Preferably, the constructed MS-MRF filter expression is as follows:

[0013]

[0014] in, This represents the expression for MS-MRF in the space-spectral joint domain; ρ represents the linear resolution ratio between the two grids; q s q represents a section function over a spatial domain; r The section function representing the spectral domain; This represents the i-th pixel in the spatial domain; Represents the i-th pixel in the spectral domain; h s and h r These represent the bandwidth used in the spatial domain and the spectral domain, respectively; x represents the input data; i represents the i-th pixel; and n represents the total number of pixels.

[0015] Preferably, the method for image filtering includes: constructing an interpretable MS-MRF convolutional filter based on a combination of prior guidance and convolutional neural networks, integrating the mathematical interpretability of the MS-MRF filter into the convolution operation in deep learning, and designing a 3×3 convolutional kernel IMMC.

[0016]

[0017] Among them, w k The total number of kernels used represents the weights of the convolution kernels; k represents the total number of kernels used internally. This represents the result of applying the MS-MRF filtering process to the input data x; K represents the total number of kernels used in the convolution operation.

[0018] Preferably, the method for posterior interpretation of the filtered image includes: in a deep learning framework, applying the IMMC convolutional kernel to the bottleneck module of the ResNet network to replace the original convolutional kernel and form a new IMMB residual block; the formed IMMB residual block achieves posterior interpretability by visualizing the activation function of the intermediate layer.

[0019]

[0020] Where A(x) represents the output activation graph of the IMMB module; x represents the input data.

[0021] The present invention also provides an interpretability system for synthetic aperture radar image filtering, the system being used to implement the above method, comprising: a partitioning module, a filtering module, and an interpretation module;

[0022] The segmentation module is used to acquire synthetic aperture radar images for initial segmentation;

[0023] The filtering module is used to perform image filtering based on the initially divided structure;

[0024] The interpretation module is used to perform posterior interpretation on the filtered image, thereby enhancing the understanding of the image.

[0025] Preferably, the workflow of the partitioning module includes: based on the prior guidance of traditional filtering algorithms, combining the mean drift algorithm with Markov random fields to construct a mean drift-Markov random field filter in the joint space-spectral domain; the MS-MRF filter performs region partitioning of the image through an upward search strategy of correlation density peaks to achieve denoising and detail preservation.

[0026] Preferably, the constructed MS-MRF filter expression is as follows:

[0027]

[0028] in, This represents the expression for MS-MRF in the space-spectral joint domain; ρ represents the linear resolution ratio between the two grids; q s q represents a section function over a spatial domain; r The section function representing the spectral domain; This represents the i-th pixel in the spatial domain; Represents the i-th pixel in the spectral domain; h s and h r These represent the bandwidth used in the spatial domain and the spectral domain, respectively; x represents the input data; i represents the i-th pixel; and n represents the total number of pixels.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] This invention constructs an interpretable SAR image filtering method by combining prior guidance with deep learning. It addresses the difficulty of balancing traditional denoising with detail preservation, providing robust denoising for regions of different properties while maintaining a balance in detail retention. Furthermore, it addresses the "black box" problem of traditional denoising processes by incorporating prior-guided mathematical interpretability into the convolutional operations of deep learning, thus constructing an interpretable convolutional filter. This invention improves the robustness and accuracy of downstream tasks through interpretable SAR image denoising. Attached Figure Description

[0031] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a schematic diagram of a method according to an embodiment of the present invention;

[0033] Figure 2 This is a schematic diagram illustrating the formation principle of the MS-MRF according to an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram illustrating the formation principle of an interpretable SAR residual filter according to an embodiment of the present invention.

[0035] Figure 4 This is a flowchart illustrating the overall process framework of an embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] Example 1

[0039] like Figure 1 The diagram shown is a schematic representation of the method flow in this embodiment, and the steps include:

[0040] S1. Acquire synthetic aperture radar images and perform initial segmentation.

[0041] Guided by prior knowledge from traditional filtering algorithms, a Mean Shift-Markov Random Field (MS-MRF) filter is constructed. The steps include combining the Mean Shift algorithm with Markov Random Fields (MRF) to build an MS-MRF filter in the joint spatial-spectral domain. This filter uses an upward search strategy based on correlation density peaks to segment synthetic aperture radar (SAR) images, achieving denoising and detail preservation. The Mean Shift algorithm is used for spatial clustering of images, while MRF is used for probabilistic modeling of pixels. By combining the advantages of both, the MS-MRF filter can better preserve edge and texture information while removing speckle noise.

[0042] This method performs well in both uniform and non-uniform regions of SAR images, effectively solving the problem of poor performance of traditional filtering methods in non-uniform regions. The specific steps are as follows:

[0043] Assume x i Given the scattering vector of the p-dimensional original polarization SAR intensity (or amplitude) image, we can construct a series of d-dimensional vectors x defined in the space-spectral joint domain. i (in Each vector corresponds to one pixel, where the two-dimensional spatial information existing in the spatial domain is represented by x. s The spatial domain portion of vector x represents the spectral information present in the spectral domain, denoted by x. r This is represented as the spectral domain portion of a vector. Assuming the same SAR image...

[0044] It consists of a finer resolution (FR) pixel array and a coarser resolution (CR) pixel array. Let... For the pixel unit at FR, For the pixel unit at CR, CR contains the number of FR pixels. times, Let I represent the linear relationship, where S and P denote the linear resolution between the two grids. Assuming each pixel in the spatial-spectral joint domain is independent and uniformly distributed, the relationship between pixels and cells is as follows: This represents a random vector belonging to fine resolution I, defined in the joint spatial spectral domain; Describes the spatial vector x defined in the joint spatial-spectral domain. s .

[0045] In the context of image analysis, latent variables typically refer to pixel features that cannot be directly observed in an image but can significantly influence the results of the filtering process. Assume v i Represents a latent variable, where , indicating vi Defined in the joint space-spectral domain, belonging to the fine resolution set I. And assume v i With x i The log-ratios are independent and identically distributed.

[0046] Standard variable x i latent variable v i conditional probability Defined using the following mathematical expression:

[0047]

[0048] Where, μ log express The expected value; σ represents the variance. It is the variance of this logarithmic ratio.

[0049] like Figure 2 The diagram shows the formation principle of the newly constructed MS-MRF, as well as the a priori guidance and mathematical interpretation of the MS-MRF filter:

[0050]

[0051] in, This represents the expression for MS-MRF in the space-spectral joint domain; ρ represents the linear resolution ratio between the two grids; q s q represents a section function over a spatial domain; r The section function representing the spectral domain; This represents the i-th pixel in the spatial domain; Represents the i-th pixel in the spectral domain; h s and h r These represent the bandwidth used in the spatial domain and the spectral domain, respectively; x represents the input data; i represents the i-th pixel; and n represents the total number of pixels.

[0052] S2. Based on the initial structural division, perform image filtering.

[0053] Based on a combination of prior guidance and convolutional neural networks, an interpretable MS-MRF convolutional (IMMC) filter is constructed. The steps include: incorporating the mathematical interpretability of the MS-MRF filter into convolution operations in deep learning, and designing a novel 3×3 convolutional kernel, IMMC. This kernel enhances the model's denoising and detail preservation capabilities by combining the density peak search strategy of MS-MRF with convolution operations. The design of the IMMC kernel is based on the unique imaging mechanism of SAR images, enabling it to better adapt to the denoising requirements of different regions.

[0054] like Figure 3The diagram shown illustrates the principle of forming an interpretable SAR residual filter. The mathematical formula for this process can be concisely expressed as follows:

[0055]

[0056] Among them, w k The total number of kernels used represents the weights of the convolution kernels; k represents the total number of kernels used internally. This represents the result of applying the MS-MRF filtering process to the input data x; K represents the total number of kernels used in the convolution operation.

[0057] Replace the convolutional kernels in the bottleneck layer of the standard ResNet50 with IMMB. The mathematical expression for IMMB residual filtering can be stated as:

[0058]

[0059] Here, ReLU represents the activation function.

[0060] S3. Perform a posteriori interpretation on the filtered image to enhance the understanding of the image.

[0061] In a deep learning framework, the IMMC convolutional kernel is applied to the bottleneck module of the ResNet network, replacing the original convolutional kernel to form a new IMMB residual block. IMMB inherits the advantages of ResNet residual blocks, such as ease of optimization and strong portability, and achieves posterior interpretability of the model through activation function visualization. This module demonstrates the model's balance mechanism between denoising and detail preservation by analyzing the activation maps of intermediate layers.

[0062] ResNet50 (IMMB) achieves posterior interpretability by visualizing the activation functions of its intermediate layers. This means that although the model may not have built-in interpretability during training, its behavior can be explained by analyzing its internal workings. Posterior interpretability is achieved through visualization techniques, aiming to enhance the model's interpretability by revealing its internal operations. The visualization method combines dimensionality reduction techniques to achieve simple and easily human-understandable visualizations. Furthermore, visualization can be combined with other techniques to further enhance understanding.

[0063] Post-hoc interpretability can be achieved by visualizing the activation graph of the IMMB module, which can be mathematically described through the following steps:

[0064] Activation map computation: For each input sample, compute the output activation map A(x) of the IMMB module:

[0065]

[0066] Furthermore, the importance of features and the filtering effect can be analyzed based on the visualization of features by the activation function: Activation maps are analyzed to determine which features are considered important by the model during the filtering process; the filtering effect is evaluated by comparing the activation maps before and after filtering to assess the effectiveness A(x) of the IMMB filter block in noise suppression and detail preservation. The overall process of this invention is as follows: Figure 4 As shown.

[0067] Example 2

[0068] This embodiment also provides an interpretability system for synthetic aperture radar image filtering, including: a partitioning module, a filtering module, and an interpretation module; the partitioning module is used to acquire synthetic aperture radar images and perform initial partitioning; the filtering module is used to perform image filtering based on the structure of the initial partitioning; and the interpretation module is used to perform posterior interpretation of the filtered images to enhance the understanding of the images.

[0069] The following will, in conjunction with this embodiment, explain in detail how the present invention solves technical problems in real life.

[0070] First, the images acquired by the synthetic aperture radar are divided using the segmentation module for initial segmentation.

[0071] Guided by prior knowledge from traditional filtering algorithms, a Mean-Shift-Markov Random Field (MS-MRF) filter is constructed. The steps include combining the MeanShift algorithm with Markov Random Fields (MRF) to build an MS-MRF filter in the joint spatial-spectral domain. This filter uses an upward search strategy based on correlation density peaks to segment the image into regions, achieving denoising and detail preservation. The MeanShift algorithm is used for spatial clustering of the image, while MRF is used for probabilistic modeling of pixels. By combining the advantages of both, the MS-MRF filter can better preserve edge and texture information while removing speckle noise.

[0072] This method performs well in both uniform and non-uniform regions of SAR images, effectively solving the problem of poor performance of traditional filtering methods in non-uniform regions. The specific steps are as follows:

[0073] Assume x i Given the scattering vector of the p-dimensional original polarization SAR intensity (or amplitude) image, we can construct a series of d-dimensional vectors x defined in the space-spectral joint domain. i (in Each vector corresponds to one pixel, where the two-dimensional spatial information existing in the spatial domain is represented by x. s The spatial domain portion of vector x represents the spectral information present in the spectral domain, denoted by x. rThis is represented as the spectral domain portion of a vector. Assuming the same SAR image...

[0074] It consists of a finer resolution (FR) pixel array and a coarser resolution (CR) pixel array. Let... For the pixel unit at FR, For the pixel unit at CR, CR contains the number of FR pixels. times, Let I represent the linear relationship, where S and P denote the linear resolution between the two grids. Assuming each pixel in the spatial-spectral joint domain is independent and uniformly distributed, the relationship between pixels and cells is as follows: This represents a random vector belonging to fine resolution I, defined in the joint spatial spectral domain; Describes the spatial vector x defined in the joint spatial-spectral domain. s .

[0075] In the context of image analysis, latent variables typically refer to pixel features that cannot be directly observed in an image but can significantly influence the results of the filtering process. Assume v i Represents a latent variable, where , indicating v i Defined in the joint space-spectral domain, belonging to the fine resolution set I. And assume v i With x i The log-ratios are independent and identically distributed.

[0076] Standard variable x i latent variable v i conditional probability Defined using the following mathematical expression:

[0077]

[0078] Where, μ log express The expected value; σ represents the variance. It is the variance of this logarithmic ratio.

[0079] like Figure 2 The diagram shows the formation principle of the newly constructed MS-MRF, as well as the a priori guidance and mathematical interpretation of the MS-MRF filter:

[0080]

[0081] in, This represents the expression for MS-MRF in the space-spectral joint domain; ρ represents the linear resolution ratio between the two grids; q s q represents a section function over a spatial domain; r The section function representing the spectral domain; This represents the i-th pixel in the spatial domain; Represents the i-th pixel in the spectral domain; h s and h r These represent the bandwidth used in the spatial domain and the spectral domain, respectively; x represents the input data; i represents the i-th pixel; and n represents the total number of pixels.

[0082] The filtering module then filters the image based on the initially defined structure.

[0083] Based on a combination of prior guidance and convolutional neural networks, an interpretable MS-MRF convolutional (IMMC) filter is constructed. The steps include: incorporating the mathematical interpretability of the MS-MRF filter into convolution operations in deep learning, and designing a novel 3×3 convolutional kernel, IMMC. This kernel enhances the model's denoising and detail preservation capabilities by combining the density peak search strategy of MS-MRF with convolution operations. The design of the IMMC kernel is based on the unique imaging mechanism of SAR images, enabling it to better adapt to the denoising requirements of different regions.

[0084] like Figure 3 The diagram shown illustrates the principle of forming an interpretable SAR residual filter. The mathematical formula for this process can be concisely expressed as follows:

[0085]

[0086] Among them, w k The total number of kernels used represents the weights of the convolution kernels; k represents the total number of kernels used internally. This represents the result of applying the MS-MRF filtering process to the input data x; K represents the total number of kernels used in the convolution operation.

[0087] Replace the convolutional kernels in the bottleneck layer of the standard ResNet50 with IMMB. The mathematical expression for IMMB residual filtering can be stated as:

[0088]

[0089] Here, ReLU represents the activation function.

[0090] Finally, the interpretation module performs a posteriori interpretation on the filtered image to enhance the understanding of the image.

[0091] In a deep learning framework, the IMMC convolutional kernel is applied to the bottleneck module of the ResNet network, replacing the original convolutional kernel to form a new IMMB residual block. IMMB inherits the advantages of ResNet residual blocks, such as ease of optimization and strong portability, and achieves posterior interpretability of the model through activation function visualization. This module demonstrates the model's balance mechanism between denoising and detail preservation by analyzing the activation maps of intermediate layers.

[0092] ResNet50 (IMMB) achieves posterior interpretability by visualizing the activation functions of its intermediate layers. This means that although the model may not have built-in interpretability during training, its behavior can be explained by analyzing its internal workings. Posterior interpretability is achieved through visualization techniques, aiming to enhance the model's interpretability by revealing its internal operations. The visualization method combines dimensionality reduction techniques to achieve simple and easily human-understandable visualizations. Furthermore, visualization can be combined with other techniques to further enhance understanding.

[0093] Post-hoc interpretability can be achieved by visualizing the activation graph of the IMMB module, which can be mathematically described through the following steps:

[0094] Activation map computation: For each input sample, compute the output activation map A(x) of the IMMB module:

[0095]

[0096] Furthermore, the importance of features and the filtering effect can be analyzed by visualizing the features based on the activation function: the activation map is analyzed to determine which features are considered important by the model during the filtering process; the filtering effect is evaluated by comparing the activation maps before and after filtering to assess the effectiveness A(x) of the IMMB filter block in terms of noise suppression and detail preservation.

[0097] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An interpretable method for synthetic aperture radar image filtering, characterized in that the steps include... include: Acquire synthetic aperture radar images and perform initial segmentation; The initial segmentation steps include: based on the prior guidance of traditional filtering algorithms, combining the mean shift algorithm with Markov random fields to construct a mean shift-Markov random field filter, i.e., the MS-MRF filter, in the joint space-spectral domain; the MS-MRF filter performs region segmentation of the image through an upward search strategy of correlation density peaks to achieve denoising and detail preservation. The constructed MS-MRF filter expression is as follows: in, This represents the expression for MS-MRF in the space-spectral joint domain; ρ represents the linear resolution ratio between the two grids; q s q represents a section function over a spatial domain; r The section function representing the spectral domain; This represents the i-th pixel in the spatial domain; Represents the i-th pixel in the spectral domain; h s and h r These represent the bandwidth used in the spatial domain and the spectral domain, respectively; x represents the input data; i represents the i-th pixel; n represents the total number of pixels; Based on the preliminary segmentation results, image filtering is performed; A posteriori interpretation is performed on the filtered image to enhance the understanding of the image.

2. The interpretability method for synthetic aperture radar image filtering according to claim 1, characterized in that, Image filtering methods include: constructing interpretable MS-MRF filters based on a combination of prior guidance and convolutional neural networks; integrating the mathematical interpretability of MS-MRF filters into convolution operations in deep learning; and designing a 3×3 convolutional kernel IMMC. in, Indicates the weights of the convolution kernel; This represents the result of applying the MS-MRF filtering process to the input data x; K represents the total number of kernels used in the convolution operation.

3. The interpretability method for synthetic aperture radar image filtering according to claim 2, characterized in that, Methods for posterior interpretation of filtered images include: in a deep learning framework, applying the IMMC convolutional kernel to the bottleneck module of a ResNet network to replace the original convolutional kernel and form a new IMMB residual block; the formed IMMB residual block achieves posterior interpretability by visualizing the activation function of the intermediate layer. Where A(x) represents the output activation graph of the IMMB module; x represents the input data.

4. An interpretability system for synthetic aperture radar image filtering, said system being used to implement the method according to any one of claims 1-3, characterized in that, include: The module consists of a partitioning module, a filtering module, and an interpretation module.

Images