Sar image noise utilization method, medium, and device

Abstract

This invention relates to the field of image generation technology, specifically to a method, medium, and device for utilizing SAR image noise. The method includes the following steps: dividing each SAR image display into a target-dominant region and a background region; performing multi-source noise modeling on the SAR image noise, decomposing the SAR image noise into inherent physical noise and external noise; injecting inherent physical noise into the target-dominant region and injecting external noise into the background region; constructing a feature entropy reconstruction loss function, by minimizing the feature entropy reconstruction loss function, the model actively compresses intra-class divergence and enhances inter-class differences, achieving directional reconstruction of the feature entropy structure, extending the effect of SAR image noise from the input layer to the deep representation structure. This invention solves the problems of noise having a single role, lacking region-aware design, not having controllable constraints from the feature distribution structure, and most acoustic strategies being static, thus improving the stability and generalization performance of SAR target detection and classification.

Description

Technical Field

[0001] This invention relates to the field of image generation technology, and in particular to a method, medium, and device for utilizing SAR image noise. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active microwave imaging system with advantages such as all-weather, all-day operation and cloud / fog penetration. It has been widely applied in military reconnaissance, marine monitoring, disaster assessment, and urban and transportation infrastructure monitoring. Unlike optical imaging, SAR images reflect the scattering mechanism between electromagnetic waves and ground targets, containing rich structural, textural, and polarization information. Therefore, it plays an irreplaceable role in intelligent interpretation tasks such as target detection, recognition, and classification.

[0003] However, because SAR imaging employs a coherent accumulation mechanism, the coherent superposition of echoes from scattering units generates typical speckle noise. This type of noise is characterized by multiplicative, spatial correlation, and strong randomness, manifesting as a fine "salt and pepper" texture at the pixel level. On the one hand, it masks the true scattering intensity of ground objects, reducing the contrast between the target and the background; on the other hand, it destroys the original geometric and textural structure, posing significant challenges to subsequent automatic detection and recognition algorithms.

[0004] With the widespread application of deep learning technology in SAR target detection and automatic target recognition (ATR), SAR image noise problems have exhibited new characteristics: On the one hand, convolutional neural networks can "adapt" to speckle noise to a certain extent under sufficient data and relatively fixed imaging conditions; on the other hand, when the number of samples is limited, the imaging scene is varied, or the process involves cross-platform or cross-band migration, noise exacerbates the fluctuations in feature distribution, making the network prone to learning unstable background "shortcut" features, resulting in insufficient generalization ability for complex scenes and open environments. Existing research has shown that networks trained on classic datasets such as MSTAR often experience a significant drop in recognition performance when tested on different incident angles, different backgrounds, or different sensor data. This is closely related to the changes in noise patterns and background statistical characteristics with the scene.

[0005] In summary, the existing SAR image target detection and classification methods have the following problems: 1) Single role of noise: Most methods treat noise such as speckle as only harmful interference, try to weaken it by filtering / denoising or use it as a simple data augmentation method, without systematically distinguishing the different mechanisms of action of physical inherent noise and external noise.

[0006] 2) Lack of region-aware design: Existing noise addition or denoising operations are mostly applied uniformly to the entire image or random sub-blocks, and rarely adopt differentiated noise processing strategies for target and background regions, making it difficult to take into account both the manifestation of target scattering structure and the modeling of background domain differences.

[0007] 3) Lack of controllable constraints on feature distribution structure: Many works only perform alignment or style perturbation at the input or feature statistics level, lacking explicit constraints on feature entropy structures such as intra-class aggregation and inter-class separation, thus lacking interpretability and controllability in the selection of noise intensity and type.

[0008] 4) Most noise strategies are static: The noise models and their parameters in existing methods are usually pre-set before training, making it difficult to adaptively adjust the noise type and intensity in different regions according to the response during model training. This limits the space for further improving robustness and generalization ability in complex scenarios and cross-domain conditions. Summary of the Invention

[0009] The main objective of this invention is to provide a method, medium, and device for utilizing SAR image noise. This invention aims to address the technical problem of how to retain or even extract useful scattering information contained in noise while suppressing its negative impact on representation learning and decision stability in SAR target detection and classification tasks, and to construct a noise processing and utilization mechanism that combines physical rationality and learning controllability.

[0010] To achieve the above objectives, this invention proposes a method for utilizing SAR image noise, comprising the following steps: S1. Divide each SAR image into a target-dominant region and a background region; S2. Decompose SAR image noise into physical inherent noise and external noise, model physical inherent noise and external noise separately, and define the area of ​​action and functional division of physical inherent noise and external noise. S3. Based on the division of action areas and functions of inherent physical noise and external noise, specifically: inherent physical noise is injected into the target-dominant area, and external noise is injected into the background area. S4. Based on the area of ​​action and functional division of physical inherent noise and external noise, a feature entropy reconstruction loss function is constructed. By minimizing the feature entropy reconstruction loss function, the model actively compresses intra-class divergence and enhances inter-class differences, thereby realizing the directional reconstruction of the feature entropy structure and extending the role of SAR image noise from the input layer to the deep representation structure. S5. By reading the network response during the model training process, the parameters of inherent physical noise and external noise are dynamically adjusted so that the noise intensity and type are automatically adjusted within a controllable range during the training phase.

[0011] A further improvement of the SAR image noise utilization method of the present invention is that S1 specifically includes the following steps: S101. SAR image input and normalization preprocessing; S102. Initial target-dominant region extraction based on annotation and model response: First, target-dominant region extraction is based on explicit annotation; then, target response extraction is based on pre-trained detection or segmentation models; and finally, auxiliary constraints are based on feature saliency maps. S103, Construction of the background area and its partitions.

[0012] A further improvement of the SAR image noise utilization method of the present invention is that S2 specifically includes the following steps: S201. Modeling the inherent physical noise includes the following steps: S20101. Physical meaning and forms of inherent physical noise: Physically inherent noise is defined as speckled noise generated by the random phase superposition of scattering units during SAR coherent imaging. This speckled noise has a multiplicative relationship with the scattering intensity of ground objects and is strongly coupled with the geometry of the target and the scattering mechanism. Physically inherent noise is uniformly modeled as speckles with a multiplicative gamma distribution, and this model is applied to normalized images. The function is defined as follows: ; in: Indicates pixel-by-pixel multiplication. The noise field is the inherent noise of the physical system. The image after being injected with physically inherent noise; S20102. Noise Distribution and Parameter Settings: The gamma distribution is selected as the statistical model for the inherent physical noise. Each pixel in the noise field... noise value satisfy: ; in: L For equivalent number of views; S20103, Methods for generating noise fields from inherent physical noise: For a size of The normalized image is used to generate a noise field with physically inherent noise as follows: Determine the equivalent number of views parameters based on the current training status. L In each pixel Independent sampling at location: ; Form a complete noise matrix , For one A real matrix of size, where It is the height of the image. It is the width of the image; S20104. Physical division of labor for inherent physical noise: The inherent physical noise is specified to only play a role in the target-dominant region, and the mask is only applied in the target-dominant region. Position of normalized image Perform multiplicative modulation; Specify the masking in the background area The location does not have the added physical inherent noise.

[0013] A further improvement of the SAR image noise utilization method of the present invention is that S2 specifically includes the following steps: S202. Modeling external noise, including the following steps: S20201. Physical meaning and forms of external noise: External noise is defined as noise that does not originate directly from the SAR coherent imaging mechanism. External noise is modeled as a combination of additive mixed noise and spectral shaping noise, and its effect on the image is defined as follows: ; in: The noise field is the external noise source. For images containing external noise; S20202, Additive Mixed Noise Model: Random fluctuations in background intensity are represented by a convex combination of Gaussian and gamma noise: ; in: For mixed weights, The standard deviation of the external noise. The mean of the Gaussian noise is... This represents the noise value of each pixel in the noise field. To represent a pixel in the background area The corresponding noise value, Indicates shape parameters as And the scale parameter is The gamma distribution; S20203, Spectrum Shaping Interference Noise Model: A structured noise field with spectral shaping superimposed on statistical noise. The specific steps are as follows: noise value of the image background region Perform a two-dimensional Fourier transform to obtain the noise value of the image background region. Spectrum in the frequency domain ; Constructing the spectral weighting function Set as: ; in: For the azimuth frequency axis, The distance-frequency axis; The center frequency of the stripe interference; Control the stripe width; A This is the spectral interference intensity coefficient; noise value of the image background region Spectrum in the frequency domain Weighting: ; in: Noise value of the image background region The weighted spectrum in the frequency domain; right Performing the inverse Fourier transform yields the structured noise field. : ; Therefore, the structured noise field In the spatial domain, it manifests as stripes or bands with specific directions, consistent with the morphology of external interference in the real SAR background; among which: This is the inverse Fourier transform; S20204. Combination and parameter set of external noise: The noise field of external noise Defined as: ; in: These are the structured noise weighting coefficients; Parameter set of external noise Recorded as: ; At the start of training, use a set of external noise parameters. The parameters in the module have fixed initial values; they are updated by the adaptive noise strategy module within a given interval. S20205. Functional Division of External Noise: The specification stipulates that external noise is injected only in the background region, that is, masking is only applied in the background region. The pixel position will be the noise field of external noise. Superimposed onto the image; no external noise is superimposed within the target's dominant region; Within the background region, the mask will be based on the background sub-region. Different sets of parameters for external noise are assigned to different sub-regions. This creates a spatially non-uniform background noise and style.

[0014] A further improvement to the SAR image noise utilization method of the present invention is that S3 specifically includes the following steps: S301. Inject physical inherent noise into the target dominant region, specifically including the following steps: S30101, Generate the physical inherent noise modulation factor of the target dominant region: Noise field based on the inherent noise of multiplicative physics Construct the target region modulation factor : ; in: The modulation parameters for the target region; All are one matrices of the same size; When the pixel belongs to the target area hour, ; When the pixel belongs to the target area hour, That is, no physical inherent noise modulation occurs; in, For the target area in pixels The modulation factor at the position, For in pixels The modulation effect of the inherent physical noise at a location on that location; S30102. Complete the injection of inherent physical noise in the target area: Perform pixel-wise multiplicative modulation on the normalized image: ; in: To complete the intermediate image after injecting inherent physical noise; S30103, Strength Constraints: The intensity of physical inherent noise is determined by the equivalent apparent number parameter. L Control, and L Always satisfy: ; This constraint applies to the noise field of the multiplicative physical inherent noise in each sampling. Perform a check beforehand; if the strategy module outputs... L If the range is exceeded, the noise field will be generated after truncating to the nearest boundary value according to the upper and lower bounds.

[0015] A further improvement to the SAR image noise utilization method of the present invention is that S3 specifically includes the following steps: S302. Inject external noise into the background area, specifically including the following steps: S30201. Assign external noise parameters to each background sub-region: For each background sub-region Generate regional noise parameter set Regional parameters are obtained by modifying global parameters. The sub-region perturbation is performed to obtain the mapping, which satisfies the following determination: ; in: For the sub-region perturbation term, set it according to the following formula: ; in: For the first The relative weight of the first background sub-region within the background region reflects the weight of the second background sub-region. The proportion of each background sub-region in the total background region. Background sub-region within the background area The total quantity; The disturbance amplitude vector is fixed as follows: ; A set of global parameters corresponding to external noise in the background region The six parameters of the direction; Mapping guarantees that the background sub-regions from arrive It features progressively increasing differences in background style; S30202. Generate the noise field of external noise for each sub-region: For each background sub-region Based on the regional noise parameter group Generate the corresponding number External noise field in each background sub-region Its statistical noise and structured noise are strictly generated according to the model, and are only replaced with sub-region parameters; S30203, Constructing the total noise field of the background region: The noise fields of each sub-region are superimposed to obtain the total noise of the background region. : ; in, For use in regulating the first External noise field in each background sub-region Weighting factors; S30204. Complete the injection of external noise into the background area: Add background noise only to the background area: ; in: This is the final noisy image for S3; S30205, Numerical range truncation: To ensure that the intensity of the noisy image falls within the network input domain, the following steps are taken: Truncation: ; Where: clip indicates truncation to the [0,1] interval pixel by pixel; for The image after being truncated As the network input for S4; S303. Summary of Injection Results and Division of Labor: The region-differential noise-added image is obtained through S30205. Its noise composition satisfies the following defining relationship: Target Dominant Area: ; Background area: ; in: Image values ​​after adding noise to the background area. The formula applies to all pixels. It holds true in all positions. The value of a pixel in the original image. For the first The noise field of each background sub-region is in the pixel Value at position, To control whether at pixel position The application of the first Noise in each sub-region.

[0016] A further improvement to the SAR image noise utilization method of the present invention is that S4 specifically includes the following steps: Calculate the feature center vector for each target category. And define the intra-class divergence based on Euclidean distance. The mathematical expression for intra-class divergence is: ; in: Indicates that the current batch belongs to the category c The sample set, For the sample Feature representation, For the index of the sample; Define inter-class divergence Specifically, it can be expressed as: ; in, For category The mean vector, For category The mean vector; Construct the feature entropy reconstruction loss function: ; in: With fixed weights; the loss function is reconstructed by minimizing feature entropy. The model actively compresses intra-class divergence and enhances inter-class differences, achieving directional reconstruction of the feature entropy structure and extending the effect of noise from the input layer to the deep representation structure. This entropy constraint is jointly optimized with the task loss to form the total loss. ; in: For fixed weights, The task loss is characterized by This is the total loss function.

[0017] A further improvement to the SAR image noise utilization method of the present invention is that S5 specifically includes the following steps: Configure a lightweight policy network Its input is the feature statistics of the current batch, including: the intra-class divergence of the features. Inter-class divergence Mission losses The change in the value and the moving average; the policy network outputs a parameter update vector, including the number of looks for physically inherent noise. L and the standard deviation of external noise Spectral interference intensity coefficient A and structured noise weighting coefficients The output of the policy network and the noise parameters are specified to have a linear incremental relationship, i.e.: ; in: For a fixed learning rate, For the updated noise parameters, These are the noise parameters before the update.

[0018] In addition, the present invention provides a readable storage medium storing a computer program adapted to be loaded by a processor and executed as described above for the SAR image noise utilization method.

[0019] In addition, the present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, runs the SAR image noise utilization method as described above.

[0020] The technical solution of the present invention has the following beneficial effects: The SAR image noise utilization method of the present invention solves the problems of single noise role, lack of regional awareness design, lack of controllable constraints from feature distribution structure and mostly static acoustic strategies by introducing a region-aware noise injection strategy based on the division of labor between physically inherent noise (PIN) and external noise (EEN) and feature entropy reconstruction constraints, thereby improving the stability and generalization performance of SAR target detection and classification.

[0021] This invention provides a clear and reproducible statistical model and parameter range for physically inherent noise (PIN) / external noise (EEN) by modeling multi-source noise, and strictly defines the effective area and functional division of the two types of noise.

[0022] In this invention, Physical Inherent Noise (PIN) plays the role of "scattering structure manifestation and texture enhancement" in the target area, while Exogenous Noise (EEN) plays the role of "background statistical modulation and domain difference construction" in the background area. The two work together in a spatial division of labor to directly weaken background shortcuts and enhance the learnability of target textures from the input space.

[0023] This invention achieves "noise-driven feature space reconstruction," a feat unattainable by existing technologies, through a feature entropy reconstruction constraint mechanism. The physically inherent noise (PIN) injected into the target region enhances the saliency of previously low-contrast or speckled scattering structures, directly reducing random fluctuations in intra-class representations. Exogenous noise (EEN) injected into the background region constructs domain style differences within the spatial partition, preventing the model from relying on fixed background textures for discrimination. This encourages the model to strengthen the discrimination dimension based on the target structure, increasing inter-class distance. Ultimately, this invention, through this entropy reconstruction constraint, enables the feature space to exhibit a stable form of "compact intra-class and separated inter-class," significantly improving recognition stability and generalization ability in complex scenes and under cross-domain conditions.

[0024] This invention achieves closed-loop control of noise injection through an adaptive noise strategy learning module, ensuring that both physical inherent noise (PIN) and external noise (EEN) remain "moderately effective" throughout the training process. This avoids both insufficient noise leading to limited feature manifestation and excessive noise damaging the target geometry or causing training instability. Unlike existing technologies that use only static, globally consistent noise enhancement methods, the noise injection behavior of this invention dynamically adjusts with the evolution of network features, thereby maximizing the positive benefits of noise in feature entropy reconstruction, robustness improvement, and cross-domain generalization, further enhancing the overall performance advantages of the system. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0026] Figure 1 This is a flowchart of the SAR image noise utilization method of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0028] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0029] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0030] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0031] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0032] This invention is applicable to various intelligent interpretation tasks based on SAR images, including but not limited to: SAR image target detection, target recognition and target classification; automatic target recognition in military intelligence reconnaissance and battlefield surveillance; marine monitoring (such as ship detection and maritime traffic surveillance), shoreline and port monitoring; urban and transportation infrastructure monitoring (such as airport runway, bridge and road target detection); remote sensing recognition tasks in high-noise and complex environments such as snow, ice, polar and sea ice monitoring; and automatic detection and classification of key targets and areas in emergency remote sensing scenarios for disasters such as earthquakes, floods, and landslides.

[0033] SAR image noise is not entirely "harmful." Under certain physical constraints and intensity conditions, it can play a positive role in downstream tasks by revealing useful scattering structures, suppressing ineffective degrees of freedom, and reshaping the feature space. Based on this core understanding, this invention constructs a five-step overall process: SAR image region-aware segmentation, multi-source noise modeling and PIN / EEN division of labor, region-differentiated noise injection strategy, feature entropy reconstruction and representation constraint mechanism, and adaptive noise strategy learning. Through these steps, this invention, for the first time, proposes a complete SAR image noise utilization technology solution from an integrated perspective of "noise availability + multi-source noise division of labor + region awareness + feature entropy reconstruction," rather than being limited to traditional denoising or simple data augmentation. It forms a complete noise utilization method from the input image space to the feature space and then to the policy space, realizing a shift from "passively suppressing noise" to "actively designing and utilizing noise."

[0034] like Figure 1 As shown, this invention proposes a method for utilizing SAR image noise, comprising the following steps: S1, SAR image region sensing and segmentation module and its implementation: To address the problem in existing technologies that "the lack of differentiated noise design between the target and background regions makes it difficult to visualize the low-visibility scattering structure of the target and the model tends to rely on background shortcut features," this invention first constructs a region-aware partitioning mechanism at the input level, dividing each SAR image display into a target-dominant region and a background region, providing spatial constraints for subsequent PIN / EEN differentiated noise injection.

[0035] S101. SAR image input and normalization preprocessing; Acquire the raw SAR image to be processed I The image can be a single-polarization, dual-polarization, or multi-polarization amplitude / intensity map. This invention performs uniform preprocessing on the image to ensure that subsequent region segmentation and noise modeling operate within a comparable grayscale range. Specifically, this includes: Perform logarithmic transformation and range normalization on the original intensity image: ; in: For normalized images, To prevent the small constants from overflowing in logarithmic operations, Norm is a linear normalization operator that maps pixel values ​​to a preset range.

[0036] After normalization, the image size is uniformly scaled or cropped.

[0037] S102. Initial target dominant region extraction based on annotation and model response: This invention defines the target-dominant region as: a region that spatially contains the main body of the target's scattered energy and exhibits a significant task-related response. To improve applicability, this step supports multiple implementation methods, which can be adaptively combined to achieve different data conditions.

[0038] 1) Target region extraction based on explicit annotation: When the training data includes annotations such as target bounding boxes, segmentation masks, or keypoints, the target dominant region mask can be directly constructed using these annotations. Specifically, it includes: ① Map the labeled target bounding box or segmentation mask to the normalized image coordinate system; ② Perform appropriate expansion or closing operations on the initial target area to make the mask include the scattering extension area around the target, thus obtaining the expanded target mask; ③ Overlay or perform a logical "OR" operation on the masks of multiple targets to form a mask of the target-dominant region of the entire image.

[0039] 2) Target response extraction based on pre-trained detection / segmentation models: When some samples lack explicit annotations, or when target regions need to be generated during the inference stage, this invention can utilize pre-trained object detection or semantic segmentation networks to obtain candidate target regions, specifically including: ① Use a detector / segmenter trained on source domain or existing labeled data to perform forward inference on the input SAR image to obtain candidate target boxes, target probability maps or pixel-level category probability maps; ② Apply confidence thresholding and non-maximum suppression (NMS) to the candidate results to remove candidate boxes with low confidence and high overlap; ③ The selected candidate boxes or high-confidence regions are converted into binary masks, and then moderately dilated and morphologically smoothed to form the target dominant region mask.

[0040] This method can obtain reliable target regions without a large amount of manual annotation and can be continuously iterated and updated in semi-supervised or self-training scenarios.

[0041] 3) Auxiliary constraints based on feature saliency maps: To further enhance the robustness of target region segmentation, this invention can also utilize the feature saliency map, attention map, or Grad-CAM (gradient-weighted class activation mapping) response map of the intermediate layers of the backbone network to provide auxiliary constraints on the target dominant region.

[0042] ① Extract intermediate layer features from the backbone network during training, aggregate their channel dimensions, and obtain a spatial saliency map; ② Apply threshold segmentation to the saliency map to extract high-response regions; ③ Perform logical AND or OR operations between the high-response region and the target region obtained from the annotation / detector to correct or expand it.

[0043] S103. Construction of the background area and its partitions: After obtaining the target dominant region mask, this invention defines the background region mask as follows: ; in: This represents a full-size mask of the same size as the image. The background area mainly includes non-target scattering, ground background, sea clutter, and urban clutter.

[0044] The present invention further divides the background area into several sub-region masks.

[0045] S2. Multi-source noise modeling and PIN / EEN division of labor: Normalized SAR images are obtained in S1. and its target dominant region mask and background area mask Subsequently, this invention performs multi-source modeling of SAR image noise, decomposing the noise into physically inherent noise (PIN) and external noise (EEN), and provides definite statistical models and parameter sets for each, which are used for subsequent regional differential noise injection.

[0046] S201. Modeling the inherent physical noise includes the following steps: S20101. Physical meaning and forms of inherent physical noise: Physically inherent noise is defined as speckle noise generated by the random phase superposition of scattering units during SAR coherent imaging. This speckle noise has a multiplicative relationship with the scattering intensity of ground objects and is strongly coupled with the geometry and scattering mechanism of the target. Physically inherent noise is uniformly modeled as speckles with a multiplicative gamma distribution, and this is applied to normalized images. The function is defined as follows: ; in: Indicates pixel-by-pixel multiplication. The noise field is the inherent noise of the physical system. The image after being injected with physically inherent noise; S20102. Noise Distribution and Parameter Settings: This invention selects the gamma distribution as the statistical model for physically inherent noise, where each pixel in the noise field... noise value satisfy:: ; in: L It is the equivalent number of views parameter and the only free parameter controlling the variance of the speckle;

[0047] S20103, Methods for generating noise fields from inherent physical noise: For a size of The normalized image is used to generate a noise field with physically inherent noise as follows: Determine the equivalent number of views parameters based on the current training status. L (Initially 1.0, subsequently updated by the strategy module), in each pixel Independent sampling at location: ; Form a complete noise matrix , For one A real matrix of size, where It is the height of the image. This is the width of the image. Specifically, this matrix represents the position of each pixel in the image (total...). The noise values ​​correspond to each pixel. It contains the noise value at each pixel location in the image, ultimately forming a complete noise matrix. It is used for noise reduction of images; At this point, the mean of the physical inherent noise (PIN) is strictly 1, and the variance is 1 / L This ensures that the overall brightness of the image is not changed in a statistical sense, and only the local scattering is textured.

[0048] S20104. Physical division of labor for inherent physical noise: The inherent physical noise is specified to only play a role in the target's dominant region, and subsequent steps will only mask the target's dominant region. Position of normalized image Multiplicative modulation is used to reveal the low-visibility scattering structure of the target; Specify the masking in the background area The location does not superimpose physical inherent noise to avoid introducing texture noise that is highly similar to the target geometry in non-target areas.

[0049] S202. Modeling external noise (EEN) includes the following steps: S20201. Physical meaning and form of external noise (EEN): Exogenous noise (EEN) is defined as noise and style disturbances that do not originate directly from the SAR coherent imaging mechanism, but are introduced by environmental background (sea clutter, electromagnetic interference, meteorological effects, etc.), platform and link incoherent noise, and human interference. Exogenous noise (EEN) mainly affects background statistical characteristics and "domain style," and its coupling with the target geometry is weaker than that of physical inherent noise (PIN).

[0050] External noise is modeled as a combination of additive mixed noise and spectral shaping noise, and its effect on the image is defined as follows: ; in: The noise field is the external noise source. For images containing external noise; S20202, Additive Mixed Noise Model: The random fluctuations in background intensity are represented by a convex combination of Gaussian noise and gamma noise: ; in: The weighting is mixed and fixed at 0.5. The standard deviation of the external noise. The mean of the Gaussian noise is usually 0, and the expected value of the Gaussian noise is denoted by . Let be the noise value of each pixel in the noise field. The noise value of the image after noise processing. Indicates shape parameters as Scale parameters are The gamma distribution; specifically, representing a pixel in the background region. The corresponding noise value. It is generated by a convex combination of Gaussian noise and gamma noise, reflecting the random fluctuations of background noise; Gaussian part: =0, The initial value is 0.05, and adjustments are only allowed within the range of [0.01, 0.1]. Gamma component: shape parameter Scale parameter =0.02, and the overall mean is forced to be 0 by subtracting the mean; The above parameters are only fine-tuned within a limited range through the policy module during training, without exceeding the limits.

[0051] S20203, Spectrum Shaping Interference Noise Model: To simulate common external noise sources such as stripe interference and orientation ambiguity, a structured noise field with spectral shaping is superimposed on the statistical noise. The specific steps are as follows: noise value of the image background region Perform a two-dimensional Fourier transform (FFT) to obtain the noise value of the image background region. Spectrum in the frequency domain ; Constructing the spectral weighting function Set as: ; in: For the azimuth frequency axis, The distance-frequency axis; The center frequency of the stripe interference is fixed at 0.25 of the total bandwidth; Control the stripe width to a fixed value of 0.05 times the bandwidth; A This is the spectral interference intensity coefficient, initially set to 0.5, and can only be adjusted within the range [0,1]. () represents the natural exponential function.

[0052] noise value of the image background region Spectrum in the frequency domain Weighting: ; in: Noise value of the image background region The weighted spectrum in the frequency domain; right Performing the inverse Fourier transform yields the structured noise field. : ; Therefore, the structured noise field In the spatial domain, it manifests as stripes or bands with specific directions, consistent with the morphology of external interference in the real SAR background; among which: This is the inverse Fourier transform; S20204. Combination and parameter set of external noise: The noise field of external noise Defined as: ; in: This is the structured noise weighting coefficient, initially set to 1.0, and can only be adjusted within [0.2]. The set of parameters of external noise is denoted as for: ; At the start of training, use a set of external noise parameters. The fixed initial values ​​of each parameter in the process, In subsequent steps, the adaptive noise strategy module updates the parameters within a given interval; the fixed initial values ​​for each parameter are: The weighting is set to 0.5 (to balance the contributions of Gaussian and gamma noise). The standard deviation of the external noise is set to 0.05; The shape parameter of the gamma distribution is set to 2; The scaling parameter for the gamma distribution is set to 0.02; The spectral interference intensity coefficient is set to 0.5. This is the structured noise weighting coefficient, set to 1.0.

[0053] S20205. Functional Division of External Noise: The specification stipulates that external noise is injected only in the background region, that is, masking is only applied in the background region. The pixel position will be the noise field of external noise. Superimposed on the image; no external noise is superimposed in the target's dominant region to avoid weakening the manifestation of the target's scattering structure by the inherent physical noise; Within the background region, subsequent steps will be based on the background sub-region mask. Different sets of parameters for external noise are assigned to different sub-regions. This creates a spatially non-uniform background noise and style.

[0054] S203. Summary of unified representation and division of labor for multi-source noise: In summary, SAR image noise is decomposed into two parts with clear mathematical forms and physical divisions: 1) Physically inherent noise PIN: Model form: multiplicative Gamma speckle, with a mean of 1; Parameter: Equivalent number of views ; Area of ​​effect: Only in The target area is the dominant region for use; Function: Enhances and reveals the target scattering structure, providing a learnable statistical texture substrate.

[0055] 2) External noise EEN: Model form: Additive Gaussian-Gamma mixture statistical noise + spectral shaping stripe interference; parameter: Furthermore, each parameter has a fixed initial value and an allowable adjustment range; Area of ​​effect: Only in The background area is used and subsequently allocated to different background sub-regions; Function: Modulates background statistical characteristics and domain style to simulate noise and interference patterns across scenes and platforms, thereby improving the model's generalization ability.

[0056] Through the above modeling, this invention provides a clear and reproducible PIN / EEN statistical model and parameter range, and strictly defines the effective area and functional division of the two types of noise, providing a solid physical and mathematical foundation for the regional differentiated noise injection (S3) and feature entropy reconstruction constraints (S4) in the subsequent steps, fundamentally different from the fuzzy and single noise processing methods in the prior art.

[0057] S3, Regional Differentiated Noise Injection Strategy (Target-Dominated Region PIN Injection + Background Region EEN Injection): The target dominant region mask is obtained in S1. Background area mask and background sub-region mask set And obtain the PIN noise field in S2. With EEN noise field After determining the statistical model and parameter set, this invention performs regionally differentiated noise injection to reveal the scattering structure in the target region and construct domain style perturbations in the background region. This step includes two sub-steps: target region PIN injection and background region partitioned EEN injection, and is performed in a fixed order of target first and then background.

[0058] S301. Inject physical inherent noise into the target dominant region, specifically including the following steps: S30101, Generate the physical inherent noise modulation factor of the target dominant region: Noise field based on the inherent noise of multiplicative physics Construct the target region modulation factor : ; in: These are the modulation parameters for the target region, which control the noise intensity and impact of the target region. For matrices of the same size consisting entirely of 1s, this formula guarantees: When the pixel belongs to the target area hour, ; When the pixel belongs to the target area hour, That is, no physical inherent noise modulation occurs; in, For the target area in pixels The modulation factor at the location. Specifically, when the pixel belongs to the target region, =0 indicates that no modulation of inherently physical noise occurs at that pixel location. In other words, the target region is at the pixel... Modulation factor at position It will control the application of inherent physical noise in the target area. =0 indicates that the pixel is not affected by noise modulation. For in pixels The modulation effect of the inherent physical noise at a given location on that location. According to the formula, when a pixel belongs to the target region (i.e., When = 0), = 1 indicates that there is no modulation with inherent physical noise, and the pixels in the target area retain their original brightness and structure.

[0059] S30102. Complete the injection of inherent physical noise in the target area: Perform pixel-wise multiplicative modulation on the normalized image: ; in: To complete the intermediate image after injecting inherent physical noise; S30103, Strength Constraints: The intensity of physical inherent noise is determined by the equivalent apparent number parameter. L Control, and L Always satisfy: ; This constraint applies to the noise field of the multiplicative physical inherent noise in each sampling. Perform a check beforehand; if the strategy module outputs... L If the range is exceeded, the noise field will be generated after truncating to the nearest boundary value according to the upper and lower bounds.

[0060] By injecting PIN into the target-dominant region, the low-visibility scattering structure within the target-dominant region is statistically enhanced and made more visible, while the global brightness of the target geometry remains unchanged.

[0061] S302. Inject external noise into the background area: This invention specifies that EEN injects only into the background region, and then injects into sub-regions within the background. The value achieves spatially non-uniform injection; the target dominant region does not overlap with EEN. The specific execution is as follows: S30201. Assign external noise parameters to each background sub-region: For each background sub-region Generate regional noise parameter set ( It is based on the specific background sub-region The set of parameters formed by the corresponding noise characteristics. The noise parameters of each sub-region can be based on... Adjustments are made to ensure that the noise style and intensity of each sub-region match the global parameter settings; region parameters are obtained by adjusting the global parameters. The sub-region perturbation is performed to obtain the mapping, which satisfies the following determination: ; in: For the sub-region perturbation term, set it according to the following formula: ; in: For the first The relative weight of the first background sub-region within the background region reflects the weight of the second background sub-region. The proportion of each background sub-region in the total background region. Background sub-region within the background area The total quantity; The disturbance amplitude vector is fixed as follows: ; A set of global parameters corresponding to external noise in the background region The six parameters of the direction, Applies to all background areas, but can be modified via... The parameters were adjusted to refine the noise characteristics of each sub-region; Mapping guarantees that the background sub-regions from arrive It features progressively increasing differences in background style; S30202. Generate the noise field of external noise for each sub-region: For each background sub-region Based on the regional noise parameter group Generate the corresponding number External noise field in each background sub-region Its statistical noise and structured noise are strictly generated according to the model, and are only replaced with sub-region parameters; S30203, Constructing the total noise field of the background region: The noise fields of each sub-region are superimposed to obtain the total noise of the background region. : ; in, For use in regulating the first External noise field in each background sub-region Weighting factors; because{ The above formula ensures that each pixel in the background region receives only the noise component of its own sub-region, provided that each pixel is a subset of the background region mask and does not overlap with any other pixel. } indicates that it is used to regulate the first Exogenous noise in each background sub-region The set of weighted factors.

[0062] S30204. Complete the injection of external noise into the background area: Add background noise only to the background area: ; in: This is the final noisy image for S3; S30205, Numerical range truncation: To ensure that the intensity of the noisy image falls within the network input domain, the following steps are taken: Truncation: ; Where: clip indicates truncation to the [0,1] interval pixel by pixel; for The image after being truncated As the network input for S4; S303. Summary of Injection Results and Division of Labor: The region-differential noise-added image is obtained through S30205. Its noise composition satisfies the following defining relationship: Target Dominant Area: ; Background area: ; in: Image values ​​after adding noise to the background area. The formula applies to all pixels. It holds true in all positions. The value of a pixel in the original image. For the first The noise field of each background sub-region is in the pixel Value at position, To control whether at pixel position The application of the first Noise in each sub-region.

[0063] Therefore, PIN performs the function of "scattering structure manifestation and texture enhancement" in the target region, while EEN performs the function of "background statistical modulation and domain difference construction" in the background region. The two work together in a spatial division of labor to directly weaken the background shortcut and enhance the learnability of the target texture from the input space, providing a stable structured input for the subsequent feature entropy reconstruction constraint (S4).

[0064] S4. Feature Entropy Reconstruction Constraint Mechanism: After completing the region-differentiated PIN / EEN injection, this invention applies the noise-added image... Input into a deep neural network and extract the high-dimensional feature representation of the penultimate layer. This feature space is used for subsequent entropy structure reconstruction. This invention argues that the positive value of noise lies not only in the explicit representation of the scattering structure or style perturbation by the input layer, but more importantly, in its ability to redistribute and modulate representational degrees of freedom in the feature space, thereby allowing the network to focus more on stable feature dimensions related to the target scattering structure. Therefore, this invention constructs explicit and optimizable "feature entropy reconstruction constraints" in the feature domain to achieve automated enhancements in intra-class aggregation and inter-class separation.

[0065] S4 specifically includes the following steps: Calculate the feature center vector for each target category. And define the intra-class divergence based on Euclidean distance. The mathematical expression for intra-class divergence is: ; in: Indicates that the current batch belongs to the category c The larger this term is, the more severe the intra-class feature dispersion. The effect of this invention in manifesting the target scattering structure using PIN will be enhanced by optimizing this term, making the intra-class distribution more compact. For the sample Feature representation, For the index of the sample; Define inter-class divergence To measure the separability between different category centers, specifically expressed as: ; in, For category The mean vector, For category The mean vector; When EEN introduces cross-scene style perturbations in the background region, the model is forced to reduce its dependence on background texture shortcuts, thereby making the inter-class structure, which is realistically dependent on target geometry and scattering, more significant, and thus improving the inter-class distance. The goal of this invention is to continuously improve this inter-class divergence through the optimization process, making the decision boundaries of different classes in the feature space clearer and more stable.

[0066] Combining the above two metrics, a feature entropy reconstruction loss function is constructed: ; in: For fixed weights, =1.0, =1.0, used to ensure that intra-class compactness and inter-class separation are strengthened simultaneously; the loss function is reconstructed by minimizing feature entropy. The model actively compresses intra-class divergence and enhances inter-class differences, achieving directional reconstruction of the feature entropy structure and extending the effect of noise from the input layer to the deep representation structure. This entropy constraint is jointly optimized with the task loss to form the total loss. ;

[0067] in: A fixed weight of 0.2 ensures that entropy structure optimization works stably without compromising task performance. The task loss is characterized by This is the total loss function.

[0068] S5. By reading the network response during the model training process, the parameters of inherent physical noise and external noise are dynamically adjusted so that the noise intensity and type are automatically adjusted within a controllable range during the training phase.

[0069] Through this mechanism, this invention achieves "noise-driven feature space reconstruction," which is unattainable by existing technologies: the PIN injected into the target region enhances the saliency of scattering structures that were originally low-contrast or masked by speckle, directly reducing random fluctuations in intra-class representation; the EEN injected into the background region constructs domain style differences within the spatial partition, preventing the model from relying on fixed background textures for discrimination, thereby prompting the model to strengthen the discrimination dimension based on the target structure and increase the inter-class distance. Ultimately, through this entropy reconstruction constraint, this invention enables the feature space to present a stable form of "compact intra-class and separated inter-class," significantly improving recognition stability and generalization ability under complex scenes and cross-domain conditions.

[0070] S5 Adaptive Noise Strategy Learning Module: In the aforementioned steps, the type, intensity, and regional distribution of PIN and EEN have been fixed. However, since the model continuously updates its feature distribution during training, its sensitivity to noise and effective range also change with the training phase. If fixed noise parameters are always used, insufficient noise may occur in the early stages of training, while excessive noise may occur after convergence, thereby reducing the overall optimization efficiency. To address this, this invention further constructs an adaptive noise policy learning module. By reading the network response during training, it dynamically adjusts the parameters of PIN and EEN, allowing the noise intensity and type to automatically adjust within a controllable range during the training phase, thus forming a closed-loop optimization mechanism of "noise injection—feature feedback—policy update".

[0071] Configure a lightweight policy network Its input is the feature statistics of the current batch, including: the intra-class divergence of the features. Inter-class divergence Mission losses The changes and moving averages can be directly obtained from the calculation of S4 without additional inference overhead. The policy network outputs a parameter update vector, including the number of looks with physical inherent noise. L and the standard deviation of external noise Spectral interference intensity coefficient A and structured noise weighting coefficients The output of the policy network and the noise parameters are specified to have a linear incremental relationship, i.e.: ; in: For a fixed learning rate, For the updated noise parameters, The noise parameters are as follows: All noise parameters are immediately truncated after the update to ensure that the PIN's look number parameter always falls within [0.5, 3.0], while the EEN's noise parameter is strictly kept within the valid range set by S2.

[0072] This adaptive update has a clear logic: when the intra-class divergence of the model was trained in the early stages... Larger inter-class divergence With a smaller target noise level, the policy network tends to reduce the number of looks at the target (PIN) (enhancing the noise intensity in the target region) and increase the perturbation coefficient of the EEN, allowing the network to rapidly strengthen the target scattering features and weaken the background shortcuts under noisy conditions. However, as training approaches convergence, intra-class divergence decreases significantly, and inter-class divergence gradually increases, the policy network output automatically reduces noise intensity, enabling the model to stably approach the optimal decision boundary under low-noise conditions. Furthermore, when the task loss... When oscillations increase or gradient directions change repeatedly, the policy network will briefly increase the perturbation amplitude of the background EEN, exposing the model to cross-domain background changes again, thereby improving generalization ability.

[0073] Through this adaptive strategy module, this invention achieves closed-loop control of noise injection, ensuring that PIN and EEN remain "moderately effective" throughout the training process. This avoids both insufficient noise leading to limited feature manifestation and excessive noise damaging the target geometry or causing training instability. Unlike existing technologies that use only static, globally consistent noise enhancement methods, the noise injection behavior of this invention dynamically adjusts with the evolution of network features, thereby maximizing the positive benefits of noise in feature entropy reconstruction, robustness improvement, and cross-domain generalization, further enhancing the overall performance advantages of the system.

[0074] The present invention also provides a readable storage medium storing a computer program adapted to be loaded by a processor and executed by any of the SAR image noise exploitation methods described above.

[0075] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, performs any of the SAR image noise utilization methods described above.

[0076] The "regional sensing noise injection and feature entropy reconstruction mechanism based on PIN / EEN division of labor" proposed in this invention has been systematically verified in multiple SAR datasets and experimental environments, including actual SAR images, public benchmark data, and simulated scenarios. Experiments cover tasks such as target detection, target classification, cross-scene recognition, and noise robustness, aiming to verify the core technical viewpoints of this invention regarding "moderate noise has a positive effect" and "regional differentiation and multi-source noise utilization can improve feature stability and generalization ability."

[0077] Experimental results show that: (1) The method of the present invention achieves significant performance improvement in multiple tasks.

[0078] In a typical SAR target classification task (10 types of ground target scenarios), compared with the no-noise model, the present invention achieves a 6%–12% improvement in classification accuracy under different noise intensity conditions; compared with the traditional "global random noise enhancement" method, the improvement is 3%–7%.

[0079] In target detection tasks, the present invention improves the mean accuracy (mAP) by approximately 4%–8% at moderate noise levels.

[0080] (2) The present invention significantly improves the stability of the model in complex noise scenarios.

[0081] Even when speckle performance is increased to 2–5 times the original, background stripe interference is added, or the background style changes across platforms, the model trained by this invention can still maintain more than 80% of the original performance, while the performance of traditional models can drop by 30%–60%. This shows that the PIN / EEN region division of labor and feature entropy reconstruction mechanism proposed in this invention can effectively reduce the model's dependence on background shortcut features and improve the physical consistency and robustness of recognition.

[0082] (3) The present invention exhibits superior generalization ability compared to the prior art under cross-domain conditions.

[0083] Under the condition of "a fixed source domain and different backgrounds / imaging parameters in the target domain", this invention improves cross-domain recognition performance by 10%–20% compared to the standard training model and by 5%–9% compared to typical style perturbation enhancement methods. The EEN spatial partitioning injection and feature entropy reconstruction proposed in this invention can effectively simulate background domain differences, enabling the model to automatically favor the target scattering structure rather than the background texture.

[0084] (4) The adaptive noise strategy module further improves training efficiency and convergence stability.

[0085] Compared to fixed noise intensity, the adaptive strategy of this invention enables the model to enter the effective learning stage faster in the early stages of training (reducing the number of training rounds by 10%–25%), while reducing learning oscillations caused by high noise in the later stages, making the overall convergence more stable.

[0086] On long training periods or large-scale SAR datasets, this mechanism can improve the performance of the final model by 2%–4%.

[0087] (5) All experiments demonstrate that the present invention has a stable and clear positive effect in realizing the functional transformation of noise "from interference to resource". The following observations were made during the experiments: 1) After injecting PIN into the target region, the response of low-visibility scattering units in the network feature layer is significantly enhanced; 2) After background injection into EEN, the inter-class distance of the network in the feature space increases significantly, while the intra-class dispersion decreases. 3) The feature entropy indicators (intra-class entropy, inter-class divergence) show a consistent trend with the final performance, indicating that the theoretical mechanism of this invention is highly consistent with the actual effect.

[0088] In marine monitoring missions, the noise division mechanism and feature entropy reconstruction mechanism of this invention are applied to analyze SAR images in different bands. Experimental results show that, compared with traditional methods, this invention can improve mAP (mean accuracy) by about 7.4% on SAR target detection datasets with different band intensities, and can still stably identify low-visibility targets under strong background clutter interference.

[0089] Experiments on SAR datasets for classifying targets such as airport runways, bridges, and roads show that the method of this invention improves classification accuracy by approximately 8.1% compared to traditional noise suppression methods.

[0090] In summary, this invention has been thoroughly verified through experiments: both the noise division mechanism and the region-aware injection strategy can stably improve performance; the feature entropy reconstruction mechanism can effectively reshape the feature space structure; and the overall system outperforms existing technologies under complex noise, cross-domain, and perturbation conditions.

[0091] The above description is only a preferred embodiment of the present invention and does not limit the scope of the present invention. All equivalent structural transformations made under the inventive concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the protection scope of the present invention.

Claims

1. A method for utilizing noise in SAR images, characterized in that, Includes the following steps: S1. Divide each SAR image into a target-dominant region and a background region; S2. Decompose SAR image noise into physical inherent noise and external noise, model physical inherent noise and external noise separately, and define the area of ​​action and functional division of physical inherent noise and external noise. S3. Based on the division of action areas and functions of inherent physical noise and external noise, specifically: inherent physical noise is injected into the target-dominant area, and external noise is injected into the background area. S4. Based on the area of ​​action and functional division of physical inherent noise and external noise, a feature entropy reconstruction loss function is constructed. By minimizing the feature entropy reconstruction loss function, the model actively compresses intra-class divergence and enhances inter-class differences, thereby realizing the directional reconstruction of the feature entropy structure and extending the role of SAR image noise from the input layer to the deep representation structure. S5. By reading the network response during the model training process, the parameters of inherent physical noise and external noise are dynamically adjusted so that the noise intensity and type are automatically adjusted within a controllable range during the training phase.

2. The SAR image noise utilization method according to claim 1, characterized in that, S1 specifically includes the following steps: S101, SAR image input and normalization preprocessing; S102. Initial target-dominant region extraction based on annotation and model response: First, target-dominant region extraction is based on explicit annotation; then, target response extraction is based on pre-trained detection or segmentation models; and finally, auxiliary constraints are based on feature saliency maps. S103, Construction of the background area and its partitions.

3. The SAR image noise utilization method according to claim 1, characterized in that, S2 specifically includes the following steps: S201. Modeling the inherent physical noise includes the following steps: S20101. Physical meaning and forms of inherent physical noise: Physically inherent noise is defined as speckled noise generated by the random phase superposition of scattering units during SAR coherent imaging. This speckled noise has a multiplicative relationship with the scattering intensity of ground objects and is strongly coupled with the geometry of the target and the scattering mechanism. Physically inherent noise is uniformly modeled as speckles with a multiplicative gamma distribution, and this model is applied to normalized images. The function is defined as follows: ； in: Indicates pixel-by-pixel multiplication. The noise field is the inherent noise of the physical system. The image after being injected with physically inherent noise; S20102. Noise Distribution and Parameter Settings: The gamma distribution is selected as the statistical model for the inherent physical noise. Each pixel in the noise field... noise value satisfy: ； in: L For equivalent number of views; S20103, Methods for generating noise fields from inherent physical noise: For a size of The normalized image is used to generate a noise field with physically inherent noise as follows: Determine the equivalent number of views parameters based on the current training status. L In each pixel Independent sampling at location: ； Form a complete noise matrix , For one A real matrix of size, where It is the height of the image. It is the width of the image; S20104. Physical division of labor for inherent physical noise: The inherent physical noise is specified to only play a role in the target-dominant region, and the mask is only applied in the target-dominant region. Position of normalized image Perform multiplicative modulation; Specify the masking in the background area The location does not have the added physical inherent noise.

4. The SAR image noise utilization method according to claim 3, characterized in that, S2 specifically includes the following steps: S202. Modeling external noise, including the following steps: S20201. Physical meaning and forms of external noise: External noise is defined as noise that does not originate directly from the SAR coherent imaging mechanism. External noise is modeled as a combination of additive mixed noise and spectral shaping noise, and its effect on the image is defined as follows: ； in: The noise field is the external noise source. For images containing external noise; S20202, Additive Mixed Noise Model: Random fluctuations in background intensity are represented by a convex combination of Gaussian and gamma noise: ； in: For mixed weights, The standard deviation of the external noise. The mean of the Gaussian noise is . This represents the noise value of each pixel in the noise field. To represent a pixel in the background area The corresponding noise value, Indicates shape parameters as And the scale parameter is The gamma distribution; S20203, Spectrum Shaping Interference Noise Model: A structured noise field with spectral shaping superimposed on statistical noise. The specific steps are as follows: noise value of the image background region Perform a two-dimensional Fourier transform to obtain the noise value of the image background region. Spectrum in the frequency domain ; Constructing the spectral weighting function Set as: ； in: For the azimuth frequency axis, The distance-frequency axis; The center frequency of the stripe interference; Control the stripe width; A This is the spectral interference intensity coefficient; noise value of the image background region Spectrum in the frequency domain Weighting: ； in: Noise value of the image background area The weighted spectrum in the frequency domain; right Performing the inverse Fourier transform yields the structured noise field. : ； Therefore, the structured noise field In the spatial domain, it manifests as stripes or bands with specific directions, consistent with the morphology of external interference in the real SAR background; among which: This is the inverse Fourier transform; S20204. Combination and parameter set of external noise: The noise field of external noise Defined as: ； in: These are the structured noise weighting coefficients; Parameter set of external noise Recorded as: ； At the start of training, use a set of external noise parameters. The parameters in the module have fixed initial values; they are updated by the adaptive noise strategy module within a given interval. S20205. Functional Division of External Noise: The specification stipulates that external noise is injected only in the background region, that is, masking is only applied in the background region. The pixel position will be the noise field of external noise. Superimposed onto the image; no external noise is superimposed within the target's dominant region; Within the background region, the mask will be based on the background sub-region. Different sets of parameters for external noise are assigned to different sub-regions. This creates a spatially non-uniform background noise and style.

5. The SAR image noise utilization method according to claim 1, characterized in that, S3 specifically includes the following steps: S301. Inject physical inherent noise into the target dominant region, specifically including the following steps: S30101, Generate the physical inherent noise modulation factor of the target dominant region: Noise field based on the inherent noise of multiplicative physics Construct the modulation factor of the target region : ； in: Masking the target dominant region; All are matrices of the same size, consisting entirely of 1s; When the pixel belongs to the target area hour, ; When the pixel belongs to the target area hour, That is, no physical inherent noise modulation occurs; in, Masking the target dominant region at the pixel The value at that location, For in pixels The modulation effect of the inherent physical noise at a location on that location; S30102. Complete the injection of inherent physical noise in the target area: Perform pixel-wise multiplicative modulation on the normalized image: ； in: To complete the intermediate image after injecting inherent physical noise; S30103, Strength Constraints: The intensity of physical inherent noise is determined by the equivalent apparent number parameter. L Control, and L Always satisfy: ； This constraint applies to the noise field of the multiplicative physical inherent noise in each sampling. Perform a check beforehand; if the strategy module outputs... L If the range is exceeded, the noise field will be generated after truncating to the nearest boundary value according to the upper and lower bounds.

6. The SAR image noise utilization method according to claim 1, characterized in that, S3 specifically includes the following steps: S302. Injecting external noise into the background area, specifically including the following steps: S30201. Assign external noise parameters to each background sub-region: For each background sub-region Generate regional noise parameter set Regional parameters are obtained by modifying global parameters. The sub-region perturbation is performed to obtain the mapping, which satisfies the following determination: ； in: For the sub-region perturbation term, set it according to the following formula: ； in: For the first The relative weight of the first background sub-region within the background region reflects the weight of the second background sub-region. The proportion of each background sub-region in the total background region. Background sub-region within the background area The total quantity; The disturbance amplitude vector is fixed as follows: ； A set of global parameters corresponding to external noise in the background region The six parameters of the direction; Mapping guarantees that the background sub-regions from arrive It features progressively increasing differences in background style; S30202. Generate the noise field of external noise for each sub-region: For each background sub-region Based on the regional noise parameter group Generate the corresponding number External noise field in each background sub-region Its statistical noise and structured noise are strictly generated according to the model, and are only replaced with sub-region parameters; S30203, Constructing the total noise field of the background region: The noise fields of each sub-region are superimposed to obtain the total noise of the background region. : ； in, For use in regulating the first External noise field in each background sub-region Weighting factors; S30204. Complete the injection of external noise into the background area: Add background noise only to the background area: ； in: This is the final noisy image of S3; S30205, Numerical range truncation: To ensure that the intensity of the noisy image falls within the network input domain, the following steps are taken: Truncation: ； Where: clip means to truncate pixel by pixel to the [0,1] interval; for The image after being truncated As the network input for S4; S303. Summary of Injection Results and Division of Labor: The region-differential noise-added image is obtained through S30205. Its noise composition satisfies the following defining relationship: Target Dominant Area: ； Background area: ； in: Image values ​​after adding noise to the background area. The formula applies to all pixels. It holds true in all positions. The value of a pixel in the original image. For the first The noise field of each background sub-region is in the pixel Value at position, To control whether at pixel position The application of the first Noise in each sub-region.

7. The SAR image noise utilization method according to claim 1, characterized in that, S4 specifically includes the following steps: Calculate the feature center vector for each target category. And define the intra-class divergence based on Euclidean distance. The mathematical expression for intra-class divergence is: ； in: Indicates that the current batch belongs to the category c The sample set, For the sample Feature representation, For the index of the sample; Define inter-class divergence Specifically, it can be expressed as: ； in, For category The mean vector, For category The mean vector; Constructing the feature entropy to reconstruct the loss function : ； in: With fixed weights; the loss function is reconstructed by minimizing feature entropy. The model actively compresses intra-class divergence and enhances inter-class differences, achieving directional reconstruction of the feature entropy structure and extending the effect of noise from the input layer to the deep representation structure. This entropy constraint is jointly optimized with the task loss to form the total loss. ； in: For fixed weights, The task loss is characterized by This is the total loss function.

8. The SAR image noise utilization method according to claim 1, characterized in that, S5 specifically includes the following steps: Configure a lightweight policy network Its input is the feature statistics of the current batch, including: the intra-class divergence of the features. Inter-class divergence Mission losses The change in the value and the moving average; the policy network outputs a parameter update vector, including the number of looks for physically inherent noise. L and the standard deviation of external noise Spectral interference intensity coefficient A and structured noise weighting coefficients The output of the policy network and the noise parameters are specified to have a linear incremental relationship, i.e.: ； in: For a fixed learning rate, For the updated noise parameters, These are the noise parameters before the update.

9. A readable storage medium, characterized in that, The readable storage medium stores a computer program adapted to be loaded by a processor and executed by the SAR image noise utilization method according to any one of claims 1-8.

10. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, which, when executed by the processor, performs the SAR image noise utilization method according to any one of claims 1-8.

Images