Synthetic aperture radar image super-resolution method and device based on deep learning

Abstract

The application provides a kind of based on deep learning's synthetic aperture radar image super-resolution method and device, belong to remote sensing science and technology field, including: with strip mode SAR original image as low-resolution source, and combine auxiliary data, obtain low-resolution image and high-resolution image;With high-resolution image input degradation network, generate and true low-resolution image distribution consistent degradation image;Low-resolution image is input scattering characteristic conversion network, generates pseudo low-resolution image;Degradation image and pseudo low-resolution image are simultaneously input scattering body decomposition super-resolution network using multi-scale branch feature decomposition structure, reconstruct high-resolution amplitude chart;Quantitative evaluation is carried out to high-resolution amplitude chart.The application can learn potential cross-domain feature association under no accurate registration condition, effectively improve the robustness of model in cross-resolution, cross-platform SAR image processing.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing science and technology, and particularly relates to a method and apparatus for super-resolution of synthetic aperture radar images based on deep learning. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active microwave remote sensing technology that continuously transmits electromagnetic waves and receives the scattered echoes from the Earth's surface via airborne or spaceborne platforms. After signal processing, it obtains high-resolution surface images that are available in all weather conditions and at all times. As remote sensing applications increasingly demand higher levels of spatial detail, traditional hardware, limited by radar aperture, orbital altitude, and acquisition modes, can no longer directly acquire higher-resolution raw data. Therefore, software-based image super-resolution (SR) technology has become an important supplement. It reconstructs high-resolution results from low-resolution images using algorithms, significantly improving the detail representation capabilities of SAR images.

[0003] Existing SAR-SR methods fall into two main categories: traditional reconstruction and deep learning. Traditional methods include three subcategories:

[0004] Interpolation methods require less computation and can run in real time, but they are prone to blurring edges and losing details.

[0005] Nonlocal mean filtering (NLM) enhances texture by utilizing image self-similarity, but it is sensitive to speckle noise, computationally intensive, and prone to artifacts.

[0006] Sparse representation reconstructs high-frequency information through dictionary learning, which has shown good results in the early stages. However, based on the linear assumption, it is difficult to characterize the complex nonlinear structure in SAR, and its performance is limited in noisy or structurally abrupt scenarios.

[0007] Deep learning has achieved breakthroughs through Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs). CNN-based works improve the resolution of natural images through end-to-end mapping, leading to improvements in remote sensing such as multi-scale convolution, edge enhancement modules, and Transformer self-attention, which enhance structural modeling. GAN structures consist of a generator and a discriminator working together. The generator outputs high-resolution images, while the discriminator supervises realism. Representative networks, such as dual-branch edge enhancement networks, lightweight LSRGANs, and improved GANs with denoising modules, are all adapted to SAR speckle noise and texture characteristics. Other research embeds physical models of point, line, and surface scatterers into the network, or uses sub-aperture decomposition to preserve phase information, attempting to further align with the SAR imaging mechanism.

[0008] However, existing technologies still face four core problems:

[0009] Data bottleneck: Deep models generally rely on paired high- and low-resolution samples for pixel-level supervised learning. However, differences in SAR satellite platforms (strip / spotting mode, orbital altitude, incident angle, polarization) lead to inconsistencies in image scale, viewing angle, and scattering characteristics. In addition, speckle noise and temporal variations make strict registration extremely difficult, and the cost of obtaining training samples is extremely high.

[0010] Migration limitations: Optical image SR methods focus on the overall structure and edges. When directly transferred to SAR, they ignore the speckle noise and scattering mechanism unique to coherent imaging, resulting in noise amplification and texture distortion.

[0011] Reconstruction contradictions: Existing networks mostly use whole-image reconstruction, which does not take into account the spatial heterogeneity of multiple scatterers mixed within SAR pixels. As a result, structural distortion, blurred details or artifacts often occur, making it difficult to simultaneously achieve texture restoration and spatial structure preservation.

[0012] Insufficient generalization: The algorithm has poor robustness in real-world remote sensing tasks with significant scale differences and complex and varied scenarios, making it difficult to meet the needs of high-precision surface monitoring.

[0013] In summary, although significant progress has been made in SAR image super-resolution, the scarcity of paired data, differences in imaging mechanisms, pixel-scale heterogeneity, and model generalization ability remain bottlenecks restricting its large-scale application. There is an urgent need for novel deep learning strategies that integrate SAR scattering characteristics with multi-mode imaging information. Summary of the Invention

[0014] To address the aforementioned technical problems, this invention proposes a deep learning-based synthetic aperture radar (SAR) image super-resolution method and apparatus. By combining a deep neural network architecture with an unpaired training strategy, it fully utilizes the scattering characteristics and texture details of SAR images, effectively improving the spatial resolution and practical application performance of the images. The specific technical solution is as follows:

[0015] A deep learning-based super-resolution method for synthetic aperture radar images includes the following steps:

[0016] Using strip mode SAR raw images as low-resolution sources and spot mode SAR raw images as high-resolution sources, and combining auxiliary data including image metadata, digital elevation model, ground cover mask, and incident angle information, coordinate system I, raster resampling and pixel intensity normalization are completed in sequence to obtain low-resolution and high-resolution images with consistent spatial reference and normalized numerical range.

[0017] The high-resolution image is input into the degradation network to generate a degradation image that is consistent with the distribution of the real low-resolution image;

[0018] The low-resolution image is input into a scattering property conversion network to generate a pseudo low-resolution image.

[0019] The degraded image and the pseudo low-resolution image are simultaneously input into the scatterer decomposition super-resolution network with a multi-scale branch feature decomposition structure. Features at different scales are extracted and fused layer by layer to reconstruct a high-resolution amplitude map.

[0020] A quantitative evaluation of the reconstructed high-resolution amplitude map was performed.

[0021] A deep learning-based synthetic aperture radar image super-resolution device includes the following modules:

[0022] The image pairing module uses the original strip mode SAR image as a low-resolution source and the original spotlight mode SAR image as a high-resolution source. It also combines auxiliary data, including image metadata, digital elevation model, ground cover mask, and incident angle information, to sequentially complete coordinate system one, raster resampling, and pixel intensity normalization, resulting in a low-resolution image and a high-resolution image with consistent spatial reference and normal numerical range.

[0023] The degradation module inputs the high-resolution image into the degradation network to generate a degradation image that is consistent with the distribution of the real low-resolution image;

[0024] The conversion module inputs the low-resolution image into the scattering characteristic conversion network to generate a pseudo low-resolution image;

[0025] The reconstruction module simultaneously inputs the degraded image and the pseudo-low-resolution image into the scatterer decomposition super-resolution network with a multi-scale branch feature decomposition structure, extracts and fuses features at different scales layer by layer, and reconstructs a high-resolution amplitude map.

[0026] The evaluation module performs a quantitative evaluation of the reconstructed high-resolution amplitude map.

[0027] An electronic device includes: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method.

[0028] A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to implement the method described thereon.

[0029] The present invention has the following beneficial effects:

[0030] This invention introduces a fusion pseudo-supervision mechanism, overcoming the high dependence of traditional algorithms on registration data. By introducing multiple cooperating deep neural network structures, it completes high-resolution image reconstruction tasks without strictly paired data. The pseudo-supervision mechanism guides the model to learn structural information and high-frequency details from unpaired high-resolution images during training by introducing multiple self-supervised signals. Compared to methods that rely entirely on real high-resolution images for supervision, this method has advantages in flexibility and adaptability, and is particularly suitable for heterogeneous SAR data from different platforms and imaging modes. Through the collaborative optimization of deep networks, the proposed method can learn potential cross-domain feature associations without precise registration, effectively improving the model's robustness and practical application capabilities in cross-resolution and cross-platform SAR image processing.

[0031] This invention introduces pixel decomposition, constructing a specialized deep learning network model to achieve effective decomposition and refined reconstruction of mixed pixels in SAR images, thereby fully restoring the detailed features and spatial layout information of different types of scatterers. This method significantly enhances the structural clarity and texture realism of super-resolution images, overcoming the limitations of existing methods in structural fidelity and texture restoration, effectively avoiding structural distortion and artifacts, and thus significantly improving the visual quality and application reliability of SAR image super-resolution processing.

[0032] This application introduces high-resolution SAR data as auxiliary information to improve the spatial resolution of medium-resolution SAR images, thereby achieving higher-quality image super-resolution reconstruction.

[0033] This invention employs a joint adversarial generative network model to fully mine and reconstruct the details and texture features in SAR images, thereby improving the image reconstruction quality. While maintaining the generation effect, it reduces the dependence on accurate registration of high and low resolution images during training, and improves the operability and adaptability of the algorithm in practical applications.

[0034] This invention utilizes the Cycle-GAN concept to perform feature degradation on high-resolution images, simulating their mapping process to low-resolution images, thereby enhancing the model's ability to learn cross-resolution feature transformations.

[0035] This invention proposes an image resolution enhancement method based on multi-scale branch feature decomposition (MSFD), which enhances the image structure representation at the detail level and generates more realistic and high-quality super-resolution images.

[0036] This invention constructs an end-to-end image super-resolution depth network model, which simplifies the traditional multi-stage process and improves the overall system's intelligence level and inference accuracy. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method of the present invention;

[0038] Figure 2 This is a schematic diagram of an image super-resolution dataset;

[0039] Figure 3 A framework diagram of an unpaired SAR image super-resolution model based on deep learning;

[0040] Figure 4 This is a comparison image of super-resolution SAR. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0042] like Figure 1 The diagram shows a flowchart of a deep learning-based synthetic aperture radar image super-resolution method according to the present invention, comprising:

[0043] Using raw SAR images in stripmap and sliding spotlight modes as input, image preprocessing is performed in conjunction with auxiliary data;

[0044] We used the pre-processed clustered image as a high-resolution image and the striped image as a low-resolution image to form a weakly paired-unregistered dataset.

[0045] Weakly paired-unregistered datasets are input into a deep learning model, high-resolution images are input into a degradation network, and the output is a degradation result consistent with the distribution of real low-resolution images. Low-resolution images are input into a scattering property conversion network to generate pseudo low-resolution images. Then, the pseudo low-resolution images and degradation results are fed into a scatterer decomposition super-resolution network with a multi-scale branch feature decomposition structure to extract and fuse features of different scales layer by layer to reconstruct a high-resolution amplitude map.

[0046] Finally, a quantitative performance evaluation of the reconstructed high-resolution amplitude map was performed, including -3 dB resolution measurements in the azimuth and range directions and calculation of the number of views (ENL), to verify the effectiveness of the method of the present invention in improving spatial resolution and suppressing noise.

[0047] The data used in this method includes: raw SAR image data (strip pattern, spotlight pattern, etc.) and auxiliary data. The auxiliary data includes image metadata, parameters required for image preprocessing, digital elevation model (DEM), ground cover mask, incident angle information, etc., which are used to support image preprocessing, sample screening, and subsequent modeling and training processes.

[0048] Raw SAR image data includes SAR images acquired in different time phases using strip and spotlight modes. The images differ in resolution, observation angle, and beam focusing method across these modes, allowing for the construction of high- and low-resolution image pairs to meet the needs of super-resolution modeling. Specifically, spotlight mode images, with their high resolution, can serve as approximate high-resolution reference images in pseudo-supervision; strip mode images, with their wide field of view and lower resolution, are used to represent objects requiring resolution enhancement.

[0049] Image metadata: Each SAR image is accompanied by complete metadata (such as a .meta.xml file), including orbital attitude information, imaging time, incident angle, heading angle, and image geometric information. This information is used for geometric correction and standardized processing during image preprocessing. Simultaneously, the viewpoint parameters (heading angle) in the metadata also serve as key input parameters for evaluating the -3 dB resolution in the azimuth and range directions, verifying the model's reconstruction capabilities and image detail enhancement effects in different directions.

[0050] Preprocessing parameters are used to ensure that multi-source SAR images have a uniform numerical and spatial reference before being input into the deep learning model. Preprocessing parameters include projection transformation parameters (such as UTM projection zone number and reference ellipsoid), resampling methods and target resolution, pixel intensity normalization parameters, filtering and cropping window sizes, etc. These parameters are used in the image preprocessing stage to perform operations such as geometric correction, raster resampling, and dynamic range normalization, thereby improving the consistency of input data and the stability of model training.

[0051] Image preprocessing: To meet the input requirements of deep learning models regarding the numerical consistency of SAR images, this invention performs standardized preprocessing on the original SAR images before model training. This step includes raster resampling under a unified projected coordinate system and normalization based on pixel distribution.

[0052] Since SAR images originate from different imaging platforms, their original coordinate systems may differ. To ensure spatial consistency of the images, this invention employs a resampling operation based on a UTM (Universal Transverse Mercator) projection system to achieve standardized and unified image spatial reference. The specific steps are as follows:

[0053] First, determine the UTM projection zone number to which the image center point belongs based on its latitude and longitude coordinates:

[0054] (1)

[0055] in, Represents the longitude of the image center point. This represents the UTM projection zone number to which it belongs.

[0056] Subsequently, the image projection is transformed to the corresponding UTM coordinate system (based on the WGS 84 ellipsoid), and raster resampling is performed using a polynomial interpolation method to ensure that the image resolution satisfies the following relationship:

[0057] (2)

[0058] in, This refers to the spatial resolution of a low-resolution image. Represented as high-resolution image resolution. The scaling factor belongs to the set of positive integers. .

[0059] SAR images typically exhibit a large dynamic range in pixel values ​​and are susceptible to speckle noise or systematic errors, which can introduce extreme values ​​and affect the training stability of deep learning models. Therefore, this invention introduces a percentile threshold cropping strategy before inputting the image into the neural network, normalizing the pixel intensity to the [-1, 1] interval. The specific calculation method is as follows:

[0060] (3)

[0061] Among them, the set of image pixel values Calculate its 0.5% and 99.5% quantiles, denoted as . and . This indicates an operation that restricts the value to the range [-1, 1]. This represents the normalized pixel value.

[0062] In summary, by standardizing UTM projection coordinates, resampling, and normalizing pixels, this invention improves the numerical stability of input data and effectively enhances the stability and generalization ability of the model during training.

[0063] Super-resolution dataset creation

[0064] This invention fully considers the differences in imaging mechanisms, spatial resolution, and observation angles of SAR images from multiple platforms and in multiple modes, and introduces the concept of "weak pairing-non-registration" to reduce the strict requirements of traditional super-resolution methods for precise pixel alignment.

[0065] During data preparation, low-resolution and high-resolution images were first constructed from the raw SAR images acquired in strip and spotlight modes, respectively. The strip mode images have a larger coverage area but lower resolution and are used as input for resolution improvement; the spotlight mode images have higher spatial resolution and can be used as input for approximate high-resolution reference images.

[0066] Before constructing training samples, this invention combines manual screening with various auxiliary data for sample quality control. The manual screening targets candidate raw SAR images and their corresponding auxiliary data (such as image metadata, DEM, incident angle information, ground cover masks, statistical features, etc.). Through a combination of manual inspection and automatic index discrimination, abnormal images with severe geometric distortion, low signal-to-noise ratio, or drastic changes in ground features are eliminated, ensuring that the retained training samples possess high signal-to-noise ratio, spatial consistency, and representativeness. The resulting data are high-quality weakly paired high- and low-resolution image pairs.

[0067] Building upon this, by applying the principles of resolution and weak pairing, an unpaired set of high- and low-resolution samples is constructed, enabling the super-resolution network to learn and train effectively without precise registration. The resulting standardized dataset is as follows: Figure 2 As shown: (a1–a8) represent high-resolution images, and (b1–b8) represent low-resolution images. (a1–a4, b1–b4) are data selected for training after manual screening, and (a5–a8, b5–b8) are data removed during the screening process. This dataset provides high-quality data support for subsequent image restoration, enhancement, and SAR basic model training. A method for quality assessment of SAR images is also discussed.

[0068] In image super-resolution technology, image quality is typically assessed using metrics such as Peak Signal-to-Noise Ratio (PSNR) or Structure Similarity Index (SSIM). However, due to the unique imaging mechanism of SAR images (strong scattering, speckle, and strong structural characteristics), optical image metrics are insufficient to accurately reflect their imaging characteristics and quality. Therefore, this invention addresses... Figure 1 Low-resolution images input into the language High-resolution images and the results generated by the super-resolution network A quantitative evaluation method that better fits the characteristics of SAR was proposed.

[0069] To better evaluate the quality and super-resolution performance of SAR images, this invention introduces two types of evaluation metrics, including a method based on -3 dB resolution and a noise intensity measurement method based on the number of looks (ENL).

[0070] Methods based on -3 dB resolution

[0071] To quantitatively evaluate the azimuth and range resolution of synthetic aperture radar (SAR) images, this invention employs a -3 dB (half-power point) width measurement method based on the imaging direction signal profile, and introduces the yaw angle from the platform's attitude metadata as a reference for direction. The yaw angle... It represents the clockwise rotation angle of the platform relative to geographic north, and can be used to determine the azimuth imaging direction under the ground projection in the image; its orthogonal direction is the range imaging direction.

[0072] Let the resolution of the pixel in the east-west (column) direction be... The resolution (meters / pixel) in the north-south (row) direction is... The increments of the azimuth and range directions relative to the pixel coordinate system are:

[0073] (4)

[0074] (5)

[0075] (6)

[0076] (7)

[0077] in, , , , These represent the pixel-by-pixel increments in the azimuth and range directions, respectively, in the row and column directions; This is the heading angle. Next, a point is selected from the SAR image. One-dimensional intensity profiles are extracted along the azimuth and range directions, and the extracted profile signals are then subjected to 10x oversampling interpolation.

[0078] (8)

[0079] (9)

[0080] in, An integer representing the pixel sequence of the profile; and These represent the intensity profiles extracted in the azimuth and range directions, respectively.

[0081] By finding the two sampling points corresponding to when the signal drops to -3 dB (half power) of its peak value, the corresponding -3 dB width can be calculated:

[0082] (10)

[0083] (11)

[0084] (12)

[0085] (13)

[0086] in, and These represent the -3 dB pixel widths in the azimuth and range directions, respectively, which are the number of pixels that are crossed when the power drops to the peak value of -3 dB on the corresponding profile. and The -3 dB ground width (in meters) represents the azimuth and range directions, respectively, and is used to characterize the actual resolution in the two directions; , and , These represent the position indices of two sampling points when the signal strength drops from the peak value to -3 dB in the azimuth and range profiles, respectively.

[0087] A method based on the number of views (ENL)

[0088] Echo count (ENL) is an important quantitative indicator for measuring the speckle noise intensity and image region homogeneity in SAR images. For a homogeneous region (e.g., open water or uniform ground), the echo intensity is statistically analyzed within a selected local window. Among them, echo intensity It is usually obtained by squared the magnitude of the complex pixel values ​​of the SAR image, i.e.

[0089] (14)

[0090] in, The signal is a complex echo signal; when the input image is stored in dB intensity form, it needs to be inversely transformed first. This is restored to a linear intensity. Based on the above echo intensity, the number of views (ENL) can be further calculated using the following formula:

[0091] (15)

[0092] Wherein, ENL represents the number of views. Represents the mathematical expectation operator. Represents the variance operator. This represents the number of pixel samples within the local window. Represents the first in the window The echo intensity of each pixel.

[0093] In summary, this invention employs a joint quality assessment method, combining a -3 dB resolution measurement method based on platform heading angle correction with a noise intensity measurement method based on the number of looks (ENL). Through a unified preprocessing workflow and directional sampling strategy, it achieves a joint quantitative analysis of the spatial resolution limit and noise suppression performance of the SAR system.

[0094] In this invention, the deep learning model adopts the US-Net (UnPair SAR Net) model, which includes a high-resolution SAR image degradation network, a low-resolution SAR scattering characteristic conversion network, and a scatterer decomposition super-resolution network, such as... Figure 3 As shown.

[0095] This invention achieves high-precision amplitude image super-resolution reconstruction without the need for HR / LR pairing through a three-level generative adversarial mechanism of "degradation simulation - scattering characteristic conversion - scatterer decomposition and reconstruction". Specifically, it first utilizes a high-resolution SAR image degradation network. For high-resolution images Degradation simulation is performed. Unlike noise degradation models using fixed kernels, US-Net models degradation as a "learnable process," enabling the hyperparameter network to adaptively capture nonlinear scattering, geometric distortion, and speckle noise in low-resolution images, achieving higher simulation accuracy and generation diversity. The specific degradation process is as follows:

[0096] (16)

[0097] in For the hyperparameters of the degenerate network, This represents an image after resolution degradation. By training adversarially against real low-resolution images, the network can learn degradation patterns that better reflect the actual conditions of SAR imaging systems. This method essentially simulates the entire process from scattering attenuation and resolution reduction to noise enhancement, achieving more complex and realistic degradation effects while preserving the interpretability of the imaging mechanism.

[0098] Secondly, to simulate the non-ideal deviations in the scattering characteristics of low-resolution SAR images, a low-resolution SAR scattering characteristic conversion network is introduced. This method is used to construct pseudo-low-resolution samples that more closely approximate the distribution of high-resolution imaging. The main function of this network is to calculate the statistical distribution bias of SAR images (such as frequency response, texture features, and target structure), rather than the changes in the ground environment itself. Therefore, this method can construct high-resolution mappings for low-resolution images, introducing scattering distribution characteristics that more closely resemble high-resolution imaging patterns while maintaining image structural consistency. The specific form is as follows:

[0099] (17)

[0100] in, Represents a true low-resolution image; This represents a low-resolution image after feature transformation. This represents an adaptive feature bias mapping extracted from the input image, used to adjust for differences in texture details, frequency distribution, and local structure, reflecting variations in imaging features under different viewpoints, target scales, or polarization configurations. By introducing this mechanism during the training phase, the network can generate pseudo-low-resolution images that more closely resemble the style of real high-resolution imaging, providing more generalizable training support for subsequent super-resolution modules.

[0101] Finally, the scatterer decomposition super-resolution network GSR simultaneously receives images with reduced resolution during the training phase. low-resolution image after feature transformation As input, a multi-scale feature decoupling structure (MSFD) is used to perform fine-grained image reconstruction. Specifically, the network processes the image... The data is decomposed into representations at different scales, with each MSFD focusing on a different scale of structural or texture feature subspace to model local variability and heterogeneity of feature responses. Finally, features at each scale are fused into a super-resolution output image.

[0102] (18)

[0103] (19)

[0104] in, These represent the inputs respectively. and Image results generated by the scatterer decomposition super-resolution network; Representing the first in the GSR network layer( ); (·) indicates the main feature extraction module. For eigenvalue decomposition operation, This indicates that features at different scales are aggregated for analysis.

[0105] To improve training stability and supervision, an auxiliary reconstruction branch is introduced at the output end of the network, utilizing... Generated super-resolution results and real high-resolution images The L1 norm is used as the primary supervision term during training, with the following input:

[0106] (20)

[0107] in, denoted by L1 norm, this loss function measures the pixel-level absolute error between the generated and real images. It measures the overall magnitude deviation between the generated and real images, exhibiting strong edge preservation and robustness to outliers, effectively guiding the network to converge to a generated result with high structural consistency.

[0108] To further enhance the realism and detail of the generated results, this invention introduces three discriminator networks during the training phase, which are respectively applied to the degradation path, transformation path, and super-resolution path. These three discriminators provide adversarial constraints to the generator by distinguishing the differences between the real and generated images, thereby continuously improving the realism and stability of the results in the "generation-discrimination" game.

[0109] Specifically:

[0110] (1) Resolution discriminator Used to distinguish pseudo-low-resolution images generated by degenerate networks. Compared to real low-resolution images To constrain the realism of the degradation simulation process, its adversarial loss is:

[0111] (twenty one)

[0112] (2) Feature discriminator Used to distinguish pseudo-low-resolution images generated by scattering property conversion networks. Compared to real low-resolution images To ensure that the spurious samples are statistically close to the true distribution, the adversarial loss is:

[0113] (twenty two)

[0114] (3) Super-resolution path discriminator Used to distinguish super-resolution images generated by super-resolution networks. Compared to true high-resolution images To enhance the texture details and visual realism of the reconstructed results, its adversarial loss is:

[0115] (twenty three)

[0116] in, This represents taking the expected values ​​of the distributions of real and generated samples, respectively. Three discriminators provide discrimination constraints from three stages: degradation path, scattering transformation, and super-resolution reconstruction, ensuring the network's realism and robustness across multiple paths and scales. Among them, the resolution discriminator... Focusing on simulating the low-resolution imaging mechanism, feature discriminator The super-discriminator focuses on characterizing the differences in scattering distribution. This directly constrains the final reconstruction result. The three elements work together to effectively avoid problems such as over-smoothing, texture distortion, and artifacts.

[0117] In summary, the total loss function of this invention consists of four parts, including: (1) degradation path adversarial loss. (1) Constrained degenerate network generates results close to the real low-resolution graph; (2) Transformation path consistency loss : Maintain the consistency of the super-resolution image structure; (3) Super-resolution path adversarial loss : Enhance the realism and high-frequency details of the generated image; (4) L1 norm constrained pixel loss In summary, the total loss function is expressed as follows:

[0118] (twenty four)

[0119] The weight of each loss item It can be adjusted according to specific task requirements to balance the focus of network training with phased goals.

[0120] Regarding parameter initialization at the start of training, this invention employs a weight initialization method based on the He Normal distribution to ensure stable gradient propagation in deep network structures. Simultaneously, to avoid excessive differences in convergence speed between different modules, all generators and discriminators use a unified AdamW optimizer, coupled with a cosine annealing learning strategy, to ensure convergence stability during training and balanced updates between the generator and discriminator.

[0121] During the inference phase, the US-Net model of this invention only requires a single low-resolution image to be super-resolutiond as input. It then sequentially passes through a scattering characteristic transformation network and a scatterer decomposition super-resolution network to quickly generate the corresponding super-resolution result. Compared to the training phase, which requires joint input of feature transformation and degradation results, the inference process is simpler and more efficient, making it suitable for direct deployment in practical remote sensing applications. The SAR super-resolution result obtained after processing by the US-Net model is shown below. Figure 4 See Table 1. Figure 4 (a1-3, b1-3, c1-3) visualizes results at different scales; (a1-3, b1-3) are the original high-resolution (spotlight) SAR visualizations; (c1-3, d1-3) are the original low-resolution (strip) SAR visualizations; (e1-3, f1-3) are the visualizations of the model results of this invention.

[0122] Table 1 Parameters of SAR Image Super-Resolution Results

[0123]

[0124] Another aspect of the present invention provides a deep learning-based synthetic aperture radar image super-resolution device, comprising the following modules:

[0125] The image pairing module uses the original strip mode SAR image as a low-resolution source and the original spotlight mode SAR image as a high-resolution source. It also combines auxiliary data, including image metadata, digital elevation model, ground cover mask, and incident angle information, to sequentially complete coordinate system one, raster resampling, and pixel intensity normalization, resulting in a low-resolution image and a high-resolution image with consistent spatial reference and normal numerical range.

[0126] The degradation module inputs the high-resolution image into the degradation network to generate a degradation image that is consistent with the distribution of the real low-resolution image;

[0127] The conversion module inputs the low-resolution image into the scattering characteristic conversion network to generate a pseudo low-resolution image;

[0128] The reconstruction module simultaneously inputs the degraded image and the pseudo-low-resolution image into the scatterer decomposition super-resolution network with a multi-scale branch feature decomposition structure, extracts and fuses features at different scales layer by layer, and reconstructs a high-resolution amplitude map.

[0129] The evaluation module performs a quantitative evaluation of the reconstructed high-resolution amplitude map.

[0130] Another aspect of the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method.

[0131] Another aspect of the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to implement the method described thereon.

[0132] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0133] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0134] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0135] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0136] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0137] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A deep learning-based super-resolution method for synthetic aperture radar images, characterized in that, Includes the following steps: Using strip mode SAR raw images as low-resolution sources and spot mode SAR raw images as high-resolution sources, and combining auxiliary data including image metadata, digital elevation model, ground cover mask, and incident angle information, coordinate system I, raster resampling and pixel intensity normalization are completed in sequence to obtain low-resolution and high-resolution images with consistent spatial reference and normalized numerical range. The high-resolution image is input into a high-resolution image degradation network to generate a degradation image that is consistent with the distribution of the real low-resolution image. The low-resolution image is input into a low-resolution scattering property conversion network to generate a pseudo low-resolution image. The degraded image and the pseudo low-resolution image are simultaneously input into the scatterer decomposition super-resolution network with a multi-scale branch feature decomposition structure. Features at different scales are extracted and fused layer by layer to reconstruct a high-resolution amplitude map. Quantitative evaluation of the reconstructed high-resolution amplitude map; High-resolution image degradation network For high-resolution images A degradation simulation was conducted, and the specific degradation process is as follows: （16） in, For the hyperparameters of the degenerate network, This represents an image with degraded resolution. Low-resolution scattering property conversion network This is used to construct pseudo-low-resolution samples that approximate high-resolution imaging distributions, specifically in the following form: （17） in, Represents a true low-resolution image; This represents a pseudo-low-resolution image after feature transformation; This represents the feature bias mapping adaptively extracted from the input image; Scatter decomposition super-resolution networks receive images with reduced resolution. Pseudo-low-resolution image after feature transformation As input, output super-resolution image: （18） （19） in, These represent the inputs respectively. and Image results generated by the scatterer decomposition super-resolution network; Representing the first in the GSR network layer( ); (·) indicates the main feature extraction module. For eigenvalue decomposition operation, This indicates that features at different scales are aggregated for analysis.

2. The deep learning-based synthetic aperture radar image super-resolution method according to claim 1, characterized in that, The coordinate system is implemented by determining the UTM projection zone based on the latitude and longitude coordinates of the image center point and completing the projection transformation. Resampling uses polynomial interpolation to maintain an integer multiple resolution relationship between low-resolution and high-resolution images.

3. The deep learning-based synthetic aperture radar image super-resolution method according to claim 1, characterized in that, Pixel intensity normalization uses percentile thresholding to map the original dynamic range to a fixed numerical range.

4. The deep learning-based synthetic aperture radar image super-resolution method according to claim 1, characterized in that, Introduce an auxiliary reconstruction branch and utilize Generated super-resolution results and real high-resolution images The L1 norm is used as the input for training, and then trained with it as a supervised term. （20） in, L1 norm is used to measure the pixel-level absolute error between the generated image and the real image.

5. A deep learning-based synthetic aperture radar image super-resolution device, used to implement the method of claim 1, characterized in that, Includes the following modules: The image pairing module uses the original strip mode SAR image as a low-resolution source and the original spotlight mode SAR image as a high-resolution source. It also combines auxiliary data, including image metadata, digital elevation model, ground cover mask, and incident angle information, to sequentially complete coordinate system one, raster resampling, and pixel intensity normalization, resulting in a low-resolution image and a high-resolution image with consistent spatial reference and normal numerical range. The degradation module inputs the high-resolution image into the degradation network to generate a degradation image that is consistent with the distribution of the real low-resolution image; The conversion module inputs the low-resolution image into the scattering characteristic conversion network to generate a pseudo low-resolution image; The reconstruction module simultaneously inputs the degraded image and the pseudo-low-resolution image into the scatterer decomposition super-resolution network with a multi-scale branch feature decomposition structure, extracts and fuses features at different scales layer by layer, and reconstructs a high-resolution amplitude map. The evaluation module performs a quantitative evaluation of the reconstructed high-resolution amplitude map.

6. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, cause the processor to implement the method described in any one of claims 1 to 4.

Images