SAR (Synthetic Aperture Radar) image denoising and super-resolution enhancement method based on Mama model

Abstract

The invention provides an SAR image denoising and super-resolution enhancement method based on a Mama model, and the method comprises the steps: carrying out the preprocessing of an input SAR original image, and obtaining a preprocessed and enhanced SAR image; performing feature extraction processing on the enhanced SAR image based on a Mama model, and fusing local convolution features and global state modeling information to obtain a fused feature image with a clear structure and complete semantics; performing generative adversarial network joint training on the fused feature image, and performing optimization through perception loss and mean square error loss to obtain a generative adversarial network; and inputting a to-be-processed image into the generator of the generative adversarial network for reasoning and reconstruction, and enhancing the detail and structure expression capability of the image to obtain a high-resolution and low-noise SAR image. According to the method, speckle noise suppression and spatial resolution improvement can be efficiently and accurately carried out on the SAR image, and the structural definition and detail expression ability of the image are remarkably enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a SAR image denoising and super-resolution enhancement method based on a Mamba model. Background Art

[0002] Synthetic Aperture Radar (SAR) is an active microwave remote sensing system capable of all-day, all-weather Earth observation. It is widely used in fields such as environmental monitoring, natural disaster assessment, resource exploration, and urban planning. However, the inherent coherent speckle noise of SAR images significantly degrades image quality, not only affecting visual quality but also posing significant challenges to subsequent image interpretation, object classification, and feature extraction. Furthermore, due to the inherent technical limitations of SAR imaging systems, the spatial resolution of SAR images is generally insufficient, making it difficult to meet the urgent demand for high-resolution imagery in sophisticated application scenarios.

[0003] Traditional SAR image denoising methods primarily include Lee, Frost, and Kuan filters, which rely on local statistical properties. These methods offer the advantages of simplicity and computational speed, but they require manual adjustment of the filter window size and related parameters for different scenarios or noise intensities. Their generalization capabilities are insufficient, making them ineffective in adapting to varying terrain types and scene variations. This often results in oversmoothing of image details or residual noise, hindering the SAR image quality from meeting the high standards required for practical applications.

[0004] In recent years, data-driven image denoising and super-resolution technologies, represented by deep learning, have made significant progress. Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs) have been gradually applied to the enhancement of remote sensing images. Although CNN and GAN models perform well in extracting local features, their limited receptive fields prevent them from effectively capturing long-range spatial correlations in images. Consequently, they struggle to accurately restore subtle structures and complex textures in SAR images, often manifesting as oversmoothing of image details or inadequate feature reconstruction. The Transformer model effectively addresses the problem of global feature extraction in images through its self-attention mechanism. However, its high computational overhead and model training complexity often hinder practical remote sensing applications due to limited computing resources and the difficulty of model optimization.

[0005] Therefore, the current remote sensing field urgently needs a high-performance SAR image enhancement method that can efficiently capture global spatial context features and has low computational complexity. Summary of the Invention

[0006] In view of this, the present invention provides a SAR image denoising and super-resolution enhancement method based on the Mamba model to solve the above problems.

[0007] The present invention provides a SAR image denoising and super-resolution enhancement method based on a Mamba model, comprising: preprocessing an input SAR original image to obtain a preprocessed and enhanced SAR image; performing feature extraction processing on the enhanced SAR image based on the Mamba model, fusing local convolution features with global state modeling information, and obtaining a fused feature image with clear structure and complete semantics; jointly training the fused feature image with a generative adversarial network, optimizing it through perceptual loss and mean square error loss, and obtaining a generative adversarial network; inputting the image to be processed into a generator of the generative adversarial network for inference reconstruction, and enhancing the image detail and structural expression capabilities, and obtaining a high-resolution, low-noise SAR image.

[0008] In another implementation of the present invention, the preprocessing of the input SAR original image to obtain a preprocessed enhanced SAR image includes: performing edge detection on the input SAR original image using a Sobel operator and a Canny operator to extract edge feature regions of the image; performing local sharpening processing on the edge feature regions using bilateral filtering and guided filtering to enhance image edge clarity; determining a low-signal region intensity threshold based on SAR image intensity characteristics to generate a low-signal region mask; performing regional differential denoising processing based on the low-signal region mask; and enhancing details of the denoised SAR image using an unsharp mask and a high-pass filter, and adjusting the image contrast to obtain a preprocessed enhanced SAR image.

[0009] In another implementation of the present invention, performing regional differentiated denoising based on the low-signal region mask includes: performing strong denoising in the low-signal region and performing weak denoising in the high-signal region based on the low-signal region mask.

[0010] In another implementation of the present invention, the enhanced SAR image is subjected to feature extraction processing based on the Mamba model, and local convolution features are fused with global state modeling information to obtain a fused feature image with clear structure and complete semantics, including: using multiple residual dense blocks to extract local features from the preprocessed SAR image, and using convolution structure and residual connection to effectively retain high-frequency information such as texture and edges in the image; based on the local feature extraction, the Mamba state space model is introduced to model the long-range dependency relationship in the image, and its causal state space architecture and continuous signal propagation mechanism are used to enhance the modeling capability of the global structure between remote pixels; local features and global state space features are feature spliced ​​to form a comprehensive feature representation, and fused and mapped through a convolution layer to obtain a fused feature image with clear structure and complete semantics.

[0011] In another implementation of the present invention, the fused feature image is jointly trained with a generative adversarial network, and optimized through perceptual loss and mean square error loss to obtain a generative adversarial network, including: taking the fused feature image as input, constructing and training a generative network, and optimizing through perceptual loss and mean square error loss; constructing and training a discriminative network, and using a U-Net structure combined with spectral normalization to improve discrimination performance; and alternately training the generator and the discriminator to complete the joint training to obtain a generative adversarial network.

[0012] In another implementation of the present invention, the image to be processed is input into the generator of the generative adversarial network for reasoning and reconstruction, and the image detail and structure expression capabilities are enhanced to obtain a high-resolution, low-noise SAR image, including: inputting the preprocessed image to be processed into the generator, performing forward reasoning, and obtaining a preliminary reconstructed image; performing detail enhancement processing on the preliminary reconstructed image, improving edge and texture information through unsharp masking technology, and outputting a high-resolution, low-noise SAR image.

[0013] In another implementation of the present invention, the method further includes: performing image quality assessment on the high-resolution, low-noise SAR image, including calculating PSNR and SSIM indicators.

[0014] Another aspect of the present invention provides a SAR image denoising and super-resolution enhancement system based on the Mamba model, comprising: a data processing module for preprocessing an input SAR original image to obtain a preprocessed enhanced SAR image; a feature fusion module for performing feature extraction processing on the enhanced SAR image based on the Mamba model, fusing local convolution features with global state modeling information to obtain a fused feature image with clear structure and complete semantics; a model training module for jointly training the fused feature image with a generative adversarial network, optimizing it through perceptual loss and mean square error loss to obtain a generative adversarial network; and a result output module for inputting the image to be processed into the generator of the generative adversarial network for inference reconstruction, and enhancing the image detail and structural expression capabilities to obtain a high-resolution, low-noise SAR image.

[0015] Another aspect of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the SAR image denoising and super-resolution enhancement method based on the Mamba model are implemented as described above.

[0016] Another aspect of the present invention provides a computer storage medium, characterized in that a computer program is stored on the computer storage medium, and when the computer program is executed by a processor, the steps of the SAR image denoising and super-resolution enhancement method based on the Mamba model as described above are implemented.

[0017] The Mamba model-based SAR image denoising and super-resolution enhancement method of the present invention efficiently models long-range dependencies in SAR images by integrating the Mamba module of the state space model. This effectively suppresses SAR image noise and significantly improves spatial resolution, significantly enhancing the image's structural clarity and detail expression capabilities. This provides technical support for the application of remote sensing images in fields such as high-precision land feature classification, disaster monitoring, and refined remote sensing interpretation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. By reading the detailed description of the embodiments below, the advantages and benefits of the solutions will become clear to those skilled in the art. The drawings are only for the purpose of illustrating preferred embodiments and are not to be considered as limiting the present invention. In the drawings:

[0019] Figure 1The figure is a flowchart of a SAR image denoising and super-resolution enhancement method based on the Mamba model according to an embodiment of the present invention.

[0020] Figure 2 The figure is a schematic diagram of the SAR image preprocessing process according to an embodiment of the present invention.

[0021] Figure 3 Schematic diagram of the MambaR-ESRGAN generator structure according to an embodiment of the present invention.

[0022] Figure 4 Schematic diagram of the MambaR-ESRGAN discriminator structure according to an embodiment of the present invention.

[0023] Figure 5 FIG2 is a schematic diagram showing the comparison effect between a reconstructed image and an original degraded image generated according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and detailedly described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in the embodiments of the present invention should fall within the scope of protection of the embodiments of the present invention.

[0025] Figure 1 A flowchart of a SAR image denoising and super-resolution enhancement method based on the Mamba model is provided in an embodiment of the present invention. Figure 1 As shown, this embodiment mainly includes:

[0026] S101 : Preprocessing the input SAR original image to obtain a preprocessed and enhanced SAR image.

[0027] S102 , performing feature extraction processing on the enhanced SAR image based on the Mamba model, fusing local convolution features with global state modeling information, and obtaining a fused feature image with clear structure and complete semantics.

[0028] S103 , performing joint training of a generative adversarial network on the fused feature image, optimizing the image through perceptual loss and mean square error loss, and obtaining a generative adversarial network.

[0029] S104: Input the image to be processed into the generator of the generative adversarial network for inference reconstruction, and enhance the image detail and structure expression capabilities to obtain a high-resolution, low-noise SAR image.

[0030] The Mamba model-based SAR image denoising and super-resolution enhancement method of the present invention efficiently models long-range dependencies in SAR images by integrating the Mamba module of the state space model. This effectively suppresses SAR image noise and significantly improves spatial resolution, significantly enhancing the image's structural clarity and detail expression capabilities. This provides technical support for the application of remote sensing images in fields such as high-precision land feature classification, disaster monitoring, and refined remote sensing interpretation.

[0031] In another implementation of the present invention, the input SAR original image is preprocessed to obtain a preprocessed enhanced SAR image, such as Figure 2 Shown, including:

[0032] The Sobel operator and the Canny operator are used to perform edge detection on the input SAR original image and extract the edge feature area of ​​the image.

[0033] For example, the specific implementation formula is as follows:

[0034]

[0035] in,

[0036]

[0037] Then, the Canny operator is used to perform non-maximum suppression and double threshold processing on G to obtain the final edge image.

[0038] Bilateral filtering and guided filtering are used to perform local sharpening processing on the edge feature area to enhance the image edge clarity.

[0039] For example, bilateral filtering and guided filtering techniques are used to perform edge sharpening enhancement on the obtained edge feature map to highlight the edge structure information in the SAR image. The specific implementation formula is as follows:

[0040] Bilateral filtering formula:

[0041]

[0042] in:

[0043] The weight function W(x,y,x′,y′) is defined as:

[0044]

[0045] The normalization function K is:

[0046]

[0047] The intensity threshold of the low-signal area is determined based on the intensity characteristics of the SAR image, and a low-signal area mask is generated.

[0048] For example, a threshold T is set according to the intensity characteristics of the SAR image, the low signal area is identified, and a low signal area mask image M(x, y) is generated. The specific definition formula is as follows:

[0049]

[0050] Perform regional differential denoising processing based on the low signal area mask.

[0051] For example, according to the mask image M(x,y), an adaptive regional difference denoising method is used for the SAR image, and the specific definition formula is as follows:

[0052]

[0053] Among them, strong denoising and weak denoising can be achieved by adjusting the parameters of non-local means filtering or selecting other algorithms suitable for SAR image denoising.

[0054] Unsharp masking and high-pass filtering are used to enhance the details of the SAR image after denoising, and the image contrast is adjusted to obtain the preprocessed enhanced SAR image.

[0055] For example, the unsharp mask technique is used to perform detail enhancement and contrast adjustment on the SAR image. The specific definition formula is as follows:

[0056] I sharp (x,y)=I denoised (x,y)+k·[I denoised (x,y)-I smooth (x,y)]

[0057] Among them, I smooth (x,y) is the denoised image I denoised The smoothed image obtained after Gaussian blurring of (x,y); k is the control factor for image detail enhancement, which usually ranges from 0.2 to 1.5.

[0058] The image preprocessing process includes steps such as edge detection, low-signal masking, adaptive denoising, and detail enhancement. Through the above processing, the detail clarity and overall contrast of the SAR image are effectively improved, laying a good foundation for the subsequent super-resolution enhancement based on the Mamba model.

[0059] In another implementation of the present invention, performing regional differentiated denoising based on the low-signal region mask includes: performing strong denoising in the low-signal region and performing weak denoising in the high-signal region based on the low-signal region mask.

[0060] In another implementation of the present invention, the feature extraction processing of the enhanced SAR image based on the Mamba model is performed, and the local convolution features are fused with the global state modeling information to obtain a fused feature image with clear structure and complete semantics, including:

[0061] Multiple residual-in-residual dense blocks (RRDB) are used to extract local features of the preprocessed SAR image, and the convolution structure and residual connection are used to effectively retain the high-frequency information such as texture and edges in the image.

[0062] For example, its basic operation can be expressed as:

[0063]

[0064] Among them, F (l-1) is the input feature of the (l-1)th layer, Represents the nonlinear mapping function of the lth RRDB block.

[0065] Based on the local feature extraction, the Mamba state space model is introduced to model the long-range dependencies in the image. Its causal state space architecture and continuous signal propagation mechanism are used to enhance the modeling ability of the global structure between remote pixels.

[0066] For example, its state update can be represented by the following simplified first-order state space model:

[0067]

[0068] y(t)=Cx(t)+Du(t)

[0069] The discretized form is:

[0070] x t =A d x t-1 +B d μ t ,y t =C d x t +D d μ t

[0071] Among them, μ t is the local feature input, x tis the state vector, y t is the output feature, A d ,B d ,C d ,D d is a discrete state space matrix. Through the state space modeling mechanism, Mamba can model large-scale spatial structures and effectively restore the contextual information destroyed by speckle noise in SAR images.

[0072] The local feature F RRDB and the global state space characteristic F Mamba Perform feature splicing to form a comprehensive feature representation, and perform fusion mapping through the convolution layer to obtain a fusion feature image F with clear structure and complete semantics out .

[0073] For example, the formula is as follows:

[0074] F out =W1·[F RRDB ,F Mamba ]+b

[0075] Among them, [F RRDB ,F Mamba ] represents the concatenation of local and global features, W1 is the fusion weight matrix, b is the bias term, and · represents the convolution operation.

[0076] In another implementation of the present invention, the performing of a generative adversarial network (GAN) joint training on the fused feature image and optimizing the GAN through perceptual loss and mean square error loss to obtain the GAN includes:

[0077] Taking the fused feature map as input, a generative network is constructed and trained, and optimized by perceptual loss and mean square error loss.

[0078] For example, the extracted fusion feature map is used as input to construct a generative network consisting of the Mamba module and the RRDB structure. By introducing a joint optimization of the mean square error loss and the perceptual loss, the generative network is trained to output a super-resolution image with rich details, thereby improving the image reconstruction accuracy and perceptual quality. The optimization objective function is defined as follows:

[0079] Mean squared error loss (MSE Loss) is used to measure the difference between the generated image and the real image at the pixel level and is defined as follows:

[0080]

[0081] Among them, n represents the total number of pixels, y i is the real image pixel value, To generate the corresponding pixel value of the image.

[0082] Perceptual Loss is used to measure the similarity of images in the feature space. It extracts multi-layer features through a pre-trained network (such as VGG) and performs weighted calculations. The expression is:

[0083]

[0084] in, represents the feature representation extracted at layer l, N l is the number of elements in the feature map of this layer, y is the real image, To generate an image.

[0085] The total loss function is a weighted combination of perceptual loss and MSE:

[0086] L G =λ1·L MSE +λ2·L Perceptual

[0087] Among them, λ1 and λ2 are hyperparameters used to control the relative weights of the two loss terms during the training process.

[0088] Construct and train a discrimination network, and use the U-Net structure combined with spectral normalization to improve the discrimination performance.

[0089] For example, a discriminant network is constructed with U-Net as the basic structure and the spectral normalization mechanism is introduced, and training is performed based on the adversarial loss function to improve the discrimination accuracy and training stability of generated images and real images.

[0090] In this step, we design a discriminator network D for image authenticity discrimination. We use a U-Net architecture to enhance its ability to discriminate multi-scale image details and global structure. To suppress gradient explosion and improve training stability, we introduce spectral normalization in each convolutional layer of the discriminator. This controls the maximum singular value of each layer's weight matrix, ensuring stable convergence during training.

[0091] The discriminant network is trained using an adversarial loss function, whose goal is to maximize the following discriminant function:

[0092] like Figure 3 As shown in Figure 2, the generator network combines the residual dense block and the Mamba module to extract local and global features, ultimately achieving high-quality super-resolution reconstruction of SAR images. Figure 4 As shown in the figure, the discriminator network uses an encoding-decoding framework, combining convolution operations with spectral normalization mechanisms to extract image features layer by layer and perform discrimination, ultimately outputting a probability map for authenticity judgment. The generator G and the discriminator D form an adversarial network, and their optimization objective adopts the standard minimum-maximum loss function form:

[0093]

[0094] Among them, x represents a real high-resolution image, which comes from the real data distribution Pdata(x); z is the input variable of the generator, which is usually a fused feature map or noise vector, sampled from the prior distribution Pz(z); G(z) represents the image result generated by the generator after receiving the input variable z; D(x) represents the judgment result of the discriminator on the input image x, that is, the probability value of it judging that the image is a real image.

[0095] During training, the discriminator maximizes this loss function to improve its ability to distinguish generated images from real images. This training step provides effective feedback for subsequent generator optimization, thus supporting the effective game mechanism of the entire generative adversarial network.

[0096] The generator and discriminator are trained alternately to complete the joint training and obtain the generative adversarial network.

[0097] For example, by alternately optimizing the generative network and the discriminative network, generative adversarial joint training is achieved to improve the authenticity of generated images and achieve stable convergence of the training system.

[0098] After completing the structural design and loss function definition of the generation network G and the identification network D, this step jointly trains the two through alternating optimization to achieve improved authenticity of the generated images and stable convergence of the training system.

[0099] During training, the generator and discriminator engage in a constant battle: on the one hand, the discriminator D learns to distinguish between real and generated images, maximizing its accuracy; on the other hand, the generator G learns to "fool" the discriminator with more realistic images, minimizing the generation loss. This entire process is based on the standard adversarial training framework and is achieved through the following optimization objectives:

[0100] G min D max V(D,G)=E x～Pdata(x) [logD(x)]+E z～Pz(z) [log(1-D(G(z)))]

[0101] Among them, the first item E x～Pdata(x) [logD(x)] represents the discriminant’s judgment result for the real image x. It is expected that its output probability D(x) will be close to 1 to ensure that the real image is accurately identified as real. The second term E z～Pz(z) [log(1-D(G(z)))| represents the discriminant's judgment result on the generated image G(z), and its output probability is expected to be close to 0, so that the forged image can be identified as a non-real image.

[0102] In this optimization framework, the discriminator D attempts to maximize the overall loss function to improve discrimination capabilities, while the generator G attempts to minimize this function, allowing its generated images to "fool" the discriminator, even if the discriminator cannot accurately distinguish between its generated images and real images. Through this adversarial mechanism, the generator gradually learns the distribution characteristics of real data, improving the structural realism and visual quality of its output images. Joint training of the fused feature maps based on a generative adversarial network (GAN) enables the coordinated optimization of the generator and discriminator, further improving the realism and detail of the reconstructed images.

[0103] In another implementation of the present invention, the image to be processed is input into the generator of the generative adversarial network for reasoning and reconstruction, and the image detail and structure expression capabilities are enhanced to obtain a high-resolution, low-noise SAR image, including: inputting the preprocessed image to be processed into the generator, performing forward reasoning, and obtaining a preliminary reconstructed image; performing detail enhancement processing on the preliminary reconstructed image, improving edge and texture information through unsharp masking technology, and outputting a high-resolution, low-noise SAR image.

[0104] For example, the preprocessed image features are input as input z to the trained generator G, and forward reasoning is performed to generate a preliminary reconstructed image

[0105] The formula is:

[0106]

[0107] Among them, z is the fused local and global feature representation, The generated super-resolution image.

[0108] This step realizes the preliminary upsampling and reconstruction of low-resolution SAR images, providing basic image information for subsequent detail enhancement.

[0109] The image spatial resolution is restored through the residual connection and upsampling module in the generator, and the edge and texture details are enhanced by combining the Unsharp Mask technology. The calculation formula is:

[0110]

[0111] Among them, I sharp (x,y) is the enhanced pixel value, To reconstruct the image initially, I blur (x,y) is a pair The result of Gaussian blur processing, k detail enhancement factor, usually takes a value between 0.2–1.5.

[0112] This step aims to enhance the edge and texture information of the image, suppress the blur effect, and further improve the structural expression ability and visual quality of the image.

[0113] In another implementation of the present invention, the method further includes: performing image quality assessment on the high-resolution, low-noise SAR image, including calculating PSNR and SSIM indicators.

[0114] For example, the image after detail enhancement is used as the final output I SR , and optionally use indicators such as PSNR and SSIM to quantitatively evaluate the quality of the generated image. In this application, PSNR and SSIM are used as examples, and no specific restrictions are placed on the evaluation indicators.

[0115] PSNR (Peak Signal-to-Noise Ratio) is used to measure the error of image reconstruction and is defined as:

[0116]

[0117] Where L is the maximum value of the pixel (usually 255), and MSE represents the mean square error between the real image and the generated image.

[0118] SSIM (Structural Similarity Index) is used to measure the similarity of images in terms of brightness, contrast, and structure. The expression is:

[0119]

[0120] Among them, μ x ,μ Y is the mean, is the variance, σ xy is the covariance, C1 and C2 are constants used to stabilize the calculation.

[0121] This step outputs the final super-resolution image and objectively evaluates the image enhancement effect through quantitative indicators to verify the actual reconstruction performance of the model.

[0122] Compared to existing image enhancement methods based on traditional filtering or convolutional neural networks, this paper introduces the Mamba state-space model and incorporates long-range dependency modeling into the image enhancement process. This effectively addresses issues such as limited receptive field, detail loss, and oversmoothing inherent in localized methods. This provides high-quality data support for applications such as high-precision object classification, disaster monitoring, and urban planning.

[0123] Through edge detection, low signal area mask construction and adaptive denoising strategy in the preprocessing stage, the present invention optimizes the denoising intensity for different signal intensity areas respectively, retaining the important edge information of ground objects; then integrates the super-resolution network structure of the Mamba module to further improve the ability to restore image spatial details. Figure 5As shown in the figure, a comparison between the original degraded image (left) and the reconstructed image (right) using our proposed method shows that our method performs better in detail restoration and noise suppression. Experimental results show that in multiple typical SAR scenarios, our proposed method significantly improves on evaluation metrics such as PSNR and SSIM compared to Real-ESRGAN.

[0124] In addition, the method proposed in this invention has good generalization ability and robustness, and is suitable for various SAR image enhancement tasks. It can provide higher-quality remote sensing image input for downstream applications such as land feature recognition, disaster monitoring, and high-precision remote sensing interpretation, and effectively promote the in-depth application of SAR remote sensing data in intelligent geographic information systems.

[0125] Another aspect of the present invention provides a SAR image denoising and super-resolution enhancement system based on the Mamba model, comprising:

[0126] Data processing module: preprocesses the input SAR original image to obtain the preprocessed enhanced SAR image.

[0127] Feature fusion module: Based on the Mamba model, feature extraction is performed on the enhanced SAR image, and local convolution features are fused with global state modeling information to obtain a fused feature image with clear structure and complete semantics.

[0128] Model training module: performing joint training of a generative adversarial network on the fused feature image, optimizing the image through perceptual loss and mean square error loss, and obtaining a generative adversarial network.

[0129] Result output module: The image to be processed is input into the generator of the generative adversarial network for reasoning and reconstruction, and the image details and structure expression capabilities are enhanced to obtain a high-resolution, low-noise SAR image.

[0130] The Mamba model-based SAR image denoising and super-resolution enhancement system of the present invention efficiently models long-range dependencies in SAR images by integrating the Mamba module of the State Space Model. This effectively suppresses SAR image noise and significantly improves spatial resolution, significantly enhancing the image's structural clarity and detail expression capabilities. This provides technical support for the application of remote sensing imagery in fields such as high-precision object classification, disaster monitoring, and refined remote sensing interpretation.

[0131] In another aspect of the present invention, an electronic device includes a processor, a memory, a communication bus, and a communication interface.

[0132] in:

[0133] The processor, memory and communication interface communicate with each other through a communication bus.

[0134] Communication interface, used to communicate with other electronic devices or servers.

[0135] The processor is used to execute the program, and specifically can execute the steps of any one of the SAR image denoising and super-resolution enhancement methods based on the Mamba model in the above embodiments.

[0136] Specifically, the program may include program codes including computer operation instructions.

[0137] The processor may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application. The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs; or different types of processors, such as one or more CPUs and one or more ASICs.

[0138] Memory is used to store programs. The memory may include high-speed RAM memory, and may also include non-volatile memory (non-volatile memory), such as at least one disk storage.

[0139] The program can be specifically configured to cause a processor to execute the steps of any of the Mamba model-based SAR image denoising and super-resolution enhancement methods described in the embodiments. The specific implementation of each step in the program can be found in the corresponding descriptions of the steps and units executed in any of the aforementioned Mamba model-based SAR image denoising and super-resolution enhancement methods, and is not further described here. Those skilled in the art will clearly understand that, for ease and brevity of description, the specific operating processes of the devices and modules described above can refer to the corresponding process descriptions in the aforementioned method embodiments.

[0140] The exemplary embodiments of the present application further provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute the methods of the various embodiments of the present application.

[0141] The method according to the embodiment of the present invention described above can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD ROM, RAM, floppy disk, hard disk or magneto-optical disk), or as computer code that is originally stored in a remote recording medium or a non-temporary machine-readable medium downloaded via a network and will be stored in a local recording medium, so that the method described herein can be stored in such software processing on a recording medium using a general-purpose computer, a dedicated processor or programmable or dedicated hardware (such as an ASIC or FPGA). It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code, and when the software or computer code is accessed and executed by a computer, a processor or hardware, the method described herein is implemented. In addition, when a general-purpose computer accesses the code for implementing the method shown here, the execution of the code converts the general-purpose computer into a dedicated computer for executing the method shown here.

[0142] Thus far, specific embodiments of the present invention have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims can be performed in a different order and still achieve the desired results. Additionally, the processes depicted in the accompanying drawings do not necessarily require the specific order shown, or sequential order, to achieve the desired results.

[0143] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, back, etc.) are only used to explain the relative position relationship between the components in a certain specific order (as shown in the accompanying drawings). If the specific order changes, the directional indication will also change accordingly.

[0144] In the description of the present invention, the terms "first" and "second" are used solely to facilitate description of different components or names and should not be construed as indicating or implying a sequential relationship, relative importance, or implicitly specifying the quantity of the technical features being described. Therefore, features specified as "first" or "second" may explicitly or implicitly include at least one of such features.

[0145] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0146] It should be noted that although the specific embodiments of the present invention are described in detail in conjunction with the accompanying drawings, this should not be construed as limiting the scope of protection of the present invention. Within the scope described by the claims, various modifications and variations that can be made by those skilled in the art without creative effort still fall within the scope of protection of the present invention.

[0147] The examples of the embodiments of the present invention are intended to briefly illustrate the technical features of the embodiments of the present invention so that those skilled in the art can intuitively understand the technical features of the embodiments of the present invention, and are not intended to improperly limit the embodiments of the present invention.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A SAR image denoising and super-resolution enhancement method based on the Mamba model, characterized in that: include: Preprocessing the input SAR original image to obtain a preprocessed enhanced SAR image; Performing feature extraction processing on the enhanced SAR image based on the Mamba model, fusing local convolution features with global state modeling information, and obtaining a fused feature image with clear structure and complete semantics; Performing joint training of a generative adversarial network on the fused feature image, optimizing the image through perceptual loss and mean square error loss to obtain a generative adversarial network; The image to be processed is input into the generator of the generative adversarial network for reasoning and reconstruction, and the image detail and structure expression capabilities are enhanced to obtain a high-resolution, low-noise SAR image.

2. The method according to claim 1, characterized in that The preprocessing of the input SAR original image to obtain the preprocessed enhanced SAR image includes: Use Sobel operator and Canny operator to perform edge detection on the input SAR original image and extract the edge feature area of ​​the image; Using bilateral filtering and guided filtering to perform local sharpening processing on the edge feature area to enhance the image edge clarity; Determine the intensity threshold of the low-signal area based on the SAR image intensity characteristics and generate a low-signal area mask; Performing regional differential denoising processing based on the low signal area mask; Unsharp masking and high-pass filtering are used to enhance the details of the SAR image after denoising, and the image contrast is adjusted to obtain the preprocessed enhanced SAR image.

3. The method according to claim 2, characterized in that The performing of regional differentiated denoising processing based on the low signal area mask includes: Based on the low signal area mask, strong denoising is performed in the low signal area, and weak denoising is performed in the high signal area.

4. The method according to claim 1, wherein The enhanced SAR image is subjected to feature extraction processing based on the Mamba model, and local convolution features are fused with global state modeling information to obtain a fused feature image with clear structure and complete semantics, including: Multiple residual dense blocks are used to extract local features from preprocessed SAR images, and convolutional structures and residual connections are used to effectively retain high-frequency information such as texture and edges in the image. Based on the local feature extraction, the Mamba state space model is introduced to model the long-range dependencies in the image. Its causal state space architecture and continuous signal propagation mechanism are used to enhance the modeling ability of the global structure between remote pixels. The local features and the global state space features are spliced ​​to form a comprehensive feature representation, and then fused and mapped through the convolution layer to obtain a fused feature image with clear structure and complete semantics.

5. The method according to claim 1, wherein The generative adversarial network is jointly trained on the fused feature image, and optimized by perceptual loss and mean square error loss to obtain a generative adversarial network, including: Taking the fused feature map as input, constructing and training a generative network, and optimizing it through perceptual loss and mean square error loss; Construct and train a discrimination network, using a U-Net structure combined with spectral normalization to improve discrimination performance; The generator and discriminator are trained alternately to complete the joint training and obtain the generative adversarial network.

6. The method according to claim 5, characterized in that The image to be processed is input into the generator of the generative adversarial network for inference reconstruction, and the image detail and structure expression capability is enhanced to obtain a high-resolution, low-noise SAR image, including: Inputting the preprocessed image to be processed into the generator, performing forward reasoning, and obtaining a preliminary reconstructed image; The preliminary reconstructed image is subjected to detail enhancement processing, and edge and texture information are enhanced by using unsharp masking technology to output a high-resolution, low-noise SAR image.

7. The method according to claim 6, characterized in that Also includes: Image quality assessment is performed on the high-resolution, low-noise SAR image, including PSNR and SSIM index calculations.

8. A SAR image denoising and super-resolution enhancement system based on the Mamba model, characterized in that: include: Data processing module: preprocesses the input SAR original image to obtain the preprocessed enhanced SAR image; Feature fusion module: performs feature extraction processing on the enhanced SAR image based on the Mamba model, fuses local convolution features with global state modeling information, and obtains a fused feature image with clear structure and complete semantics; Model training module: performing joint training of a generative adversarial network on the fused feature image, optimizing the image through perceptual loss and mean square error loss to obtain a generative adversarial network; Result output module: The image to be processed is input into the generator of the generative adversarial network for reasoning and reconstruction, and the image details and structure expression capabilities are enhanced to obtain a high-resolution, low-noise SAR image.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the SAR image denoising and super-resolution enhancement method based on the Mamba model are implemented as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that The computer storage medium stores a computer program, which, when executed by a processor, implements the steps in the SAR image denoising and super-resolution enhancement method based on the Mamba model according to any one of claims 1 to 7.

Images