Strong scattering point feature guided maritime synthetic aperture radar image generation method and device

Abstract

The invention discloses a strong scattering point feature guided marine synthetic aperture radar image generation method and device. The method comprises the following steps: acquiring an initial marine synthetic aperture radar image; preprocessing the initial maritime synthetic aperture radar image by using a multi-attribute combination control strategy to obtain a target feature vector sequence; constructing a strong scattering point set according to the initial maritime synthetic aperture radar image; constructing a loss weight matrix according to the strong scattering point set; constructing a target loss function according to the loss weight matrix; according to the target loss function, a maritime synthetic aperture radar image generation model is trained, and the maritime synthetic aperture radar image generation model comprises a generation diffusion sub-model; and generating a target maritime synthetic aperture radar image by using the trained maritime synthetic aperture radar image generation model. According to the method, marine synthetic aperture radar image generation is realized, image quality is improved, and computing resource consumption is reduced. The method can be widely applied to the technical field of image processing.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and apparatus for generating marine synthetic aperture radar images guided by strong scattering point features. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an advanced remote sensing technology that uses radar systems mounted on moving platforms such as satellites or aircraft to transmit and receive electromagnetic signals, combined with synthetic aperture processing techniques, to achieve high-resolution two-dimensional imaging for Earth observation. SAR has all-weather, all-time operating capabilities, able to penetrate clouds, smoke, and even some vegetation and ground cover. However, acquiring SAR images involves complex flight experiments, data acquisition, and signal processing procedures, requiring not only high hardware costs and long observation cycles but also precise data annotation by professionals to ensure the reliability of subsequent research. To alleviate the problem of insufficient data, existing technologies employ various SAR image simulation generation methods to expand datasets and promote the training and optimization of related algorithms. Traditional SAR image generation techniques are based on electromagnetic simulation methods, but these require extremely high geometric accuracy and material property parameters of the target model, and have high computational complexity, making it difficult to meet the needs of large-scale data generation and consuming significant computational resources. Traditional methods also directly apply optical image enhancement techniques (such as image rotation and noise addition) to SAR images. However, due to the unique physical characteristics of SAR images, such as speckle noise and geometric distortion, these methods often lead to problems such as a significant reduction in signal-to-noise ratio and distortion of shadow features, which seriously affect the quality and application value of the generated images, resulting in low image quality.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0004] The main objective of this invention is to propose a method and apparatus for generating marine synthetic aperture radar images guided by strong scattering point features, thereby achieving marine synthetic aperture radar image generation, improving image quality, and reducing computational resource consumption.

[0005] On one hand, embodiments of the present invention provide a method for generating marine synthetic aperture radar images guided by strong scattering point features, comprising the following steps:

[0006] Acquire initial synthetic aperture radar images of the sea;

[0007] The initial marine synthetic aperture radar image is preprocessed using a multi-attribute combined control strategy to obtain a target feature vector sequence;

[0008] Based on the initial marine synthetic aperture radar image, a set of strong scattering points is constructed;

[0009] Based on the set of strong scattering points, construct the loss weight matrix;

[0010] Based on the loss weight matrix, construct the target loss function;

[0011] Based on the target loss function, a marine synthetic aperture radar image generation model is trained. The marine synthetic aperture radar image generation model includes a generation-diffusion sub-model, which is used to add noise and denoise the target feature vector sequence.

[0012] The trained marine synthetic aperture radar image generation model is used to generate a target marine synthetic aperture radar image.

[0013] In some embodiments, the preprocessing of the initial marine synthetic aperture radar image using a multi-attribute combined control strategy to obtain a target feature vector sequence includes:

[0014] The initial marine synthetic aperture radar image is divided into blocks and encoded to obtain a first vector sequence;

[0015] Attribute information is extracted from the initial marine synthetic aperture radar image, including target category, polarization mode, background brightness, azimuth angle, or time step.

[0016] The attribute information is normalized.

[0017] The normalized attribute information is mapped to a high-dimensional feature space using a neural network embedding layer to obtain the attribute feature vector;

[0018] The first vector sequence is concatenated with multiple attribute feature vectors to obtain a second vector sequence;

[0019] The second vector sequence is embedded with position encoding to obtain the target feature vector sequence.

[0020] In some embodiments, constructing a set of strong scattering points based on the initial marine synthetic aperture radar image includes:

[0021] Initialize the set of strong scattering points;

[0022] The initial marine synthetic aperture radar image is subjected to mean filtering to obtain a filtered image;

[0023] Add all pixels in the filtered image to the set of strong scattering points;

[0024] The pixel with the largest value is selected as the strong scattering point from the filtered image;

[0025] A strong scattering region is generated based on the strong scattering point and the preset side length;

[0026] Delete multiple target candidate points from the set of strong scattering points, where the target candidate points correspond to the pixels in the strong scattering region;

[0027] The process continues until the number of strong scattering points in the set of strong scattering points reaches a preset number.

[0028] In some embodiments, constructing a loss weight matrix based on the set of strong scattering points includes:

[0029] The location of the strong scattering points is determined based on the set of strong scattering points.

[0030] Distortion detection processing is performed on the locations of multiple strong scattering points to obtain the noise addition time step size;

[0031] The loss weight matrix is ​​constructed based on the noise addition time step size.

[0032] In some embodiments, constructing the target loss function based on the loss weight matrix includes:

[0033] The target loss function is constructed based on the loss weight matrix, time step, target category attribute condition, azimuth attribute condition, polarization mode attribute condition, and background brightness attribute condition.

[0034] In some embodiments, the noise addition and denoising process includes forward noise addition, and the noise addition and denoising process on the target feature vector sequence includes:

[0035] Calculate the variance based on the noise variance coefficient and the identity matrix;

[0036] Calculate the mean based on the noise variance coefficient and the image from the previous step;

[0037] Calculate the conditional probability distribution based on the variance, the mean, and the current step image;

[0038] Based on the noise addition time step size and the conditional probability distribution, calculate the probability distribution from the original image to the noisy image;

[0039] The noisy image is calculated based on the probability distribution from the original image to the noisy image and the noise following a Gaussian distribution.

[0040] In some embodiments, the noise addition and denoising process includes inverse denoising processing, and the noise addition and denoising process on the target feature vector sequence includes:

[0041] Calculate the probability distribution from the noisy image to the original image;

[0042] The original image is calculated based on the probability distribution from the noisy image to the original image, the random Gaussian noise, and the weighting coefficients.

[0043] In some embodiments, the process of constructing the marine synthetic aperture radar image generation model includes:

[0044] Construct a shallow network, which includes multiple Transformer layers;

[0045] After the shallow network, a Transformer module is built;

[0046] After the Transformer module, a deep network is constructed, which makes skip connections with the shallow network. The deep network includes multiple Transformer layers.

[0047] A normalization layer is constructed after the deep network;

[0048] After the normalization layer, a convolutional layer is constructed.

[0049] On the other hand, embodiments of the present invention provide a marine synthetic aperture radar image generation apparatus guided by strong scattering point features, comprising:

[0050] The first module is used to acquire initial marine synthetic aperture radar images;

[0051] The second module is used to preprocess the initial marine synthetic aperture radar image using a multi-attribute combined control strategy to obtain a target feature vector sequence.

[0052] The third module is used to construct a set of strong scattering points based on the initial marine synthetic aperture radar image;

[0053] The fourth module is used to construct a loss weight matrix based on the set of strong scattering points;

[0054] The fifth module is used to construct the target loss function based on the loss weight matrix;

[0055] The sixth module is used to train the marine synthetic aperture radar image generation model according to the target loss function. The marine synthetic aperture radar image generation model includes a generation-diffusion sub-model, which is used to add noise and denoise the target feature vector sequence.

[0056] The seventh module is used to generate target marine synthetic aperture radar images using the trained marine synthetic aperture radar image generation model.

[0057] On the other hand, embodiments of the present invention provide a computer device, including:

[0058] At least one processor;

[0059] At least one memory for storing at least one program;

[0060] When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0061] On the other hand, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0062] On the other hand, embodiments of the present invention provide a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0063] The embodiments of this application include at least the following beneficial effects: This application provides a method and apparatus for generating marine synthetic aperture radar images guided by strong scattering point features. The embodiments of this application first acquire an initial marine synthetic aperture radar image, preprocess the initial marine synthetic aperture radar image using a multi-attribute combination control strategy to obtain a target feature vector sequence, then construct a set of strong scattering points and a loss weight matrix based on the initial marine synthetic aperture radar image, and then construct a target loss function based on the loss weight matrix to train the marine synthetic aperture radar image generation model, and finally use the trained marine synthetic aperture radar image generation model to generate the target marine synthetic aperture radar image, thereby enabling marine synthetic aperture radar image generation through strong scattering point feature guidance, improving image quality and reducing computational resource consumption.

[0064] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the description and the drawings. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a flowchart of a method for generating marine synthetic aperture radar images guided by strong scattering point features, according to an embodiment of the present invention.

[0067] Figure 2 This is a schematic diagram of an architecture for generating marine synthetic aperture radar images according to an embodiment of the present invention;

[0068] Figure 3 This is a flowchart of a multi-attribute combination control strategy according to an embodiment of the present invention;

[0069] Figure 4 This is a flowchart illustrating how to construct a loss weight matrix based on a target's strong scattering points, according to an embodiment of the present invention.

[0070] Figure 5 This is a schematic diagram of the visualization results of extracting strong scattering points from a maritime ship target according to an embodiment of the present invention;

[0071] Figure 6 This is a schematic diagram of a noise addition and denoising process according to an embodiment of the present invention;

[0072] Figure 7 This is a schematic diagram of a flowchart for obtaining various SAR target attributes according to an embodiment of the present invention;

[0073] Figure 8 This is a comparison diagram of the results of weighted and unweighted methods guided by strong scattering point features according to an embodiment of the present invention;

[0074] Figure 9 This is a comparison chart of MS-SSIM indices under different categories according to an embodiment of the present invention;

[0075] Figure 10 This is a schematic diagram of the multi-attribute combination control generation result of a SAR image according to an embodiment of the present invention;

[0076] Figure 11 This is a schematic diagram of a marine synthetic aperture radar image generation device guided by strong scattering point features according to an embodiment of the present invention.

[0077] Figure 12 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0079] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0080] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0081] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0082] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0083] Synthetic Aperture Radar (SAR) is a high-resolution imaging radar that can obtain high-resolution radar images similar to optical photography even in extremely low visibility weather conditions. It utilizes the relative motion between the radar and the target to synthesize a larger equivalent antenna aperture from smaller real antenna apertures through data processing; hence, it is also called synthetic aperture radar. The characteristics of SAR include high resolution, all-weather operation, and effective identification of camouflage and penetration of concealment.

[0084] Among related technologies, Synthetic Aperture Radar (SAR) is an advanced remote sensing technology. It transmits and receives electromagnetic signals through radar systems mounted on moving platforms such as satellites or aircraft, and combines this with synthetic aperture processing techniques to achieve high-resolution two-dimensional imaging for Earth observation. Compared to optical remote sensing, SAR has all-weather, all-time operating capabilities, able to penetrate clouds, smoke, and even some vegetation and ground cover, thus enabling stable acquisition of observational data even under complex weather conditions. Based on these advantages, SAR technology has been widely applied in fields such as topographic mapping, environmental monitoring, military reconnaissance, disaster assessment, and resource exploration. Although SAR technology possesses these advantages, the acquisition of SAR images involves complex flight experiments, data acquisition, and signal processing procedures. This not only requires high hardware costs and long observation cycles but also relies on precise data annotation by professionals to ensure the reliability of subsequent research. These factors result in a limited size of currently available SAR image datasets, severely restricting the development of applications such as SAR image interpretation and target recognition. To alleviate the problem of insufficient data, researchers have proposed a variety of SAR image simulation generation methods to expand the dataset and promote the training and optimization of related algorithms.

[0085] To generate better SAR images, it is first necessary to analyze their image features. Unlike optical images, SAR images exhibit significant feature differences due to their unique imaging method, among which strong scattering points are one of the most discriminative features. Strong scattering points are generated by specular or angular reflection effects of the target's local structure, appearing as high-brightness pixels in the image. Their formation mechanism has three key characteristics: first, the scattering intensity is significantly higher than the background area, forming obvious bright spot features; second, the spatial distribution exhibits azimuth sensitivity, showing a regular evolution with changes in radar viewing angle; and finally, in the complex domain, they appear as phase-coherent scattering centers. These characteristics make strong scattering points an important carrier of target features in SAR images, not only constituting the geometric skeleton features of the target but also providing key discrimination criteria for image interpretation. Furthermore, since maritime targets are severely affected by background factors such as clutter and sea state, the stability characteristics of target strong scattering points become a key advantage for the detection and identification of ships and other maritime targets. Therefore, in the process of SAR image generation, especially maritime SAR image generation, accurate modeling of strong scattering point features is crucial to ensuring the physical authenticity and application effectiveness of the generated images.

[0086] Furthermore, in the process of simulating and generating marine SAR images, the quality of SAR image generation is also closely related to the precise control of its feature attributes. Unlike optical images, marine SAR image data can be decomposed into multiple key attribute dimensions, including target category, azimuth, background brightness, and polarization mode. These attribute parameters together determine the application value and technical feasibility of marine SAR images. Specifically, they include: (1) Target category attribute characterizes the reflection characteristics of targets in the image. For example, different types of ship targets, such as fishing boats and cargo ships, exhibit significantly different scattering characteristics and texture structures in SAR images. Accurately generating images of specific categories is very important for downstream target recognition tasks. (2) Azimuth attribute reflects the relative horizontal orientation relationship between the radar and the target, directly affecting the geometric deformation characteristics of the target in the image. (3) Background brightness attribute mainly reflects the brightness of the sea surface, which is affected by the size of the sea waves and the radar observation angle. Specifically, when the sea surface wind speed is low, the image background will become significantly brighter as the wind speed increases; when the wind and waves are large, dark spots will appear due to wave obstruction. When simulating and modeling, it is necessary to accurately reproduce the brightness contrast relationship between the background and the target under different sea conditions in order to ensure the authenticity of the target scattering features in the generated image. (4) The polarization mode attribute reflects the polarization interaction characteristics between radar waves and the target, which directly affects the integrity of the target scattering information and the feature representation capability. Different polarization combinations can reveal different dielectric properties and geometric structural features of the target. Therefore, the generation of high-quality marine SAR images not only requires accurate modeling of the strong scattering point features of the target, but also requires the coordinated control of the above-mentioned multiple attributes. Only by accurately controlling the joint distribution of these feature parameters can we ensure that the generated marine SAR image has both physical authenticity and meets the technical requirements of actual application scenarios. Therefore, developing a marine SAR image generation method guided by the strong scattering point of the target and with joint controllable multiple attributes has important theoretical value and practical significance for promoting the application of SAR technology.

[0087] In existing technologies, traditional SAR image generation techniques are mainly based on electromagnetic simulation methods. This approach involves three key steps: first, constructing an accurate physical model of the target; second, simulating SAR echo signals using ray tracing algorithms; and finally, processing the echo data using imaging algorithms to generate SAR images. While this method can produce high-fidelity SAR images, it demands extremely high geometric accuracy and material property parameters from the target model, and its computational complexity is high, making it difficult to meet the needs of large-scale data generation. Furthermore, some studies have attempted to directly apply optical image enhancement techniques (such as image rotation and noise addition) to SAR images. However, due to the unique physical characteristics of SAR images, such as speckle noise and geometric distortion, these methods often lead to a significant reduction in signal-to-noise ratio and distortion of shadow features, severely impacting the quality and application value of the generated images. Another common method is based on image segmentation and synthesis techniques, which generates new images by segmenting and recombining real SAR images. However, this method has significant limitations: on the one hand, the segmentation process disrupts the continuous scattering characteristics of the target; on the other hand, the recombined image struggles to maintain the azimuth consistency of the original target, resulting in geometric features of the generated image that do not conform to the true SAR imaging mechanism. Therefore, while traditional SAR image generation methods have some effect on data augmentation, they generally suffer from inherent defects such as poor image quality and high computational resource consumption, making it difficult to meet the demand for high-quality and diverse SAR image data in practical applications. This technical bottleneck urgently needs to be overcome.

[0088] In recent years, the rapid development of deep learning technology has brought new breakthroughs to the field of SAR image generation. In particular, methods based on Generative Adversarial Networks (GANs) have made significant progress in this field. The GAN framework consists of two core components: a generator and a discriminator. The generator generates synthetic images by learning the data distribution of SAR images, while the discriminator is responsible for distinguishing between real and generated images. Through this adversarial training mechanism, the generator can gradually improve the realism of the generated images. In the field of SAR image generation, researchers have proposed a variety of improved GAN architectures and training strategies. For example, by introducing attention mechanisms, improving loss functions, and optimizing network structures, the quality and diversity of generated SAR images have been significantly improved. These improvements make the generated SAR images closer to real SAR images in terms of texture details and scattering features. However, existing GAN-based methods still have some inherent defects: (1) the training process is unstable and requires a lot of time to design model parameters; (2) the ability to model the features of strong scattering points of targets needs to be improved; and (3) the physical consistency of generated images needs to be improved. Meanwhile, the shortcomings of the existing technology include: (1) The existing SAR image generation technology is obviously insufficient in terms of image quality and diversity, it is difficult to accurately model the target scattering features and texture details, the model generalization ability is poor, and it cannot adapt to the changing environmental scenes and feature attribute conditions, resulting in the generation of complex environmental scenes lacking realism and completeness; (2) The current SAR image simulation technology is insufficient in modeling the scattering attribute features of multiple types of targets, and is limited by the image background environment factors, making it difficult to effectively generate images containing diverse background attributes, affecting the practical value of the marine SAR image simulation results; (3) The existing SAR image generation method focuses too much on the generation of target texture details, while ignoring the key modeling of the strong scattering point features of marine SAR targets, resulting in defects in the accurate modeling of targets in marine SAR generated images; (4) Most of the current conditional generation technologies can only achieve single controllable generation of target categories, and are insufficient in the controllable generation of key attributes such as azimuth angle and complex background, and lack a coordination control mechanism between multiple attributes, making it impossible to achieve stable marine SAR image generation with joint control of multiple attributes.

[0089] In view of this, and addressing the shortcomings and improvement needs of existing SAR image generation technologies, this invention innovatively proposes a method and apparatus for generating marine synthetic aperture radar images guided by strong scattering point features. This method, through a deep neural network architecture, enhances the modeling capability of strong scattering point features of targets and achieves precise coordinated control of multi-dimensional attributes such as target category, azimuth, and background environment, thereby improving the realism, detail richness, and physical consistency of the generated images. The attribute coupling mechanism designed in this invention effectively coordinates the interrelationships between various feature parameters, ensuring stable generation performance even in complex and ever-changing application scenarios, and providing reliable simulation data support for high-precision applications such as maritime reconnaissance and environmental monitoring.

[0090] This application provides a method for generating marine synthetic aperture radar images guided by strong scattering point features, relating to the field of image processing technology. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or vehicle terminal, but is not limited thereto. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the method for generating marine synthetic aperture radar images guided by strong scattering point features, but is not limited to the above forms.

[0091] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0092] The embodiments of this application will be explained in detail below with reference to the accompanying drawings:

[0093] Figure 1 This is an optional flowchart of a method for generating marine synthetic aperture radar images guided by strong scattering point features, provided in an embodiment of this application. Figure 1 The method may include, but is not limited to, steps S101 to S107.

[0094] Step S101: Acquire initial marine synthetic aperture radar image;

[0095] Step S102: Preprocess the initial marine synthetic aperture radar image using a multi-attribute combined control strategy to obtain a target feature vector sequence;

[0096] Step S103: Construct a set of strong scattering points based on the initial synthetic aperture radar image at sea;

[0097] Step S104: Construct the loss weight matrix based on the set of strong scattering points;

[0098] Step S105: Construct the target loss function based on the loss weight matrix;

[0099] Step S106: Train the marine synthetic aperture radar image generation model according to the target loss function. The marine synthetic aperture radar image generation model includes a generation and diffusion sub-model, which is used to add noise and denoise the target feature vector sequence.

[0100] Step S107: Generate a target marine synthetic aperture radar image using the trained marine synthetic aperture radar image generation model.

[0101] Steps S101 to S107 as shown in the embodiments of this application realize the generation of marine synthetic aperture radar images, improve image quality, and reduce computational resource consumption.

[0102] In some embodiments, the architecture for generating marine synthetic aperture radar images is as follows: Figure 2As shown, the acquired SAR image can be input into a multi-attribute conditional combination control module to preprocess the initial marine synthetic aperture radar image using a multi-attribute combination control strategy, obtaining a target feature vector sequence. This target feature vector sequence is then input into a generative diffusion sub-model constructed from multiple Transformer layers for noise addition and denoising. After layer normalization and convolutional layers, predicted noise is obtained. Finally, inverse denoising is performed to generate a target marine synthetic aperture radar image. It is understood that this embodiment designs a method for weighted calculation of loss functions on strong scattering points of the target, used to guide the generative diffusion sub-model based on the Transformer architecture to generate SAR images. While ensuring the quality of the generated image, it further processes the relevant attribute feature information (target category, azimuth, background, polarization, etc.) of the marine SAR image. The processed attribute feature information is then concatenated as a conditional input to the model network, jointly controlling the image generation process, ultimately achieving multi-attribute jointly controllable marine area SAR image generation.

[0103] In step S101 of some embodiments, an initial marine synthetic aperture radar image can be obtained through a radar database. However, the initial marine synthetic aperture radar image can also be obtained through other means, and is not limited to these.

[0104] In some embodiments, in step S102, the initial marine synthetic aperture radar image is preprocessed using a multi-attribute combined control strategy to obtain a target feature vector sequence, which may include, but is not limited to, the following steps:

[0105] The initial marine synthetic aperture radar image is divided into blocks and encoded to obtain the first vector sequence;

[0106] Attribute information is extracted from the initial marine synthetic aperture radar image. The attribute information includes target category, polarization, background brightness, azimuth angle or time step.

[0107] Normalize the attribute information;

[0108] By using the neural network embedding layer, the normalized attribute information is mapped to a high-dimensional feature space to obtain the attribute feature vector;

[0109] The second vector sequence is obtained by concatenating multiple attribute feature vectors with the first vector sequence.

[0110] Position encoding embedding is performed on the second vector sequence to obtain the target feature vector sequence.

[0111] In some embodiments, this embodiment uses Transformer as the main network architecture for training and inference. Due to Transformer's attention mechanism, it can better capture long-range dependencies between images, resulting in better model performance when processing images with complex textures and shapes, significantly improving the quality of generated SAR images. Simultaneously, this embodiment uses multiple Transformer blocks in the main network architecture, which further enhances model performance. A skip connection scheme is used between shallow and deep networks, effectively fusing shallow and deep feature information and enabling cross-layer information transfer, thereby better capturing global context information and improving model performance without incurring unnecessary computation. To better embed multiple attribute information of marine SAR images into the generation process, achieving multi-attribute combination control of marine SAR image generation and improving the accuracy and reliability of the generated results, this embodiment designs a multi-attribute combination control strategy. This strategy coordinates and controls attributes such as target category, azimuth, polarization, and background throughout the generation process, achieving multi-attribute combination control generation. The process of the multi-attribute combination control strategy is as follows: Figure 3 As shown, the initial marine synthetic aperture radar image can be first divided into blocks for encoding to obtain a first vector sequence. For example, a noisy image block can be transformed into a first vector sequence with a hidden dimension of 768, depending on the size of the given image block. Then, attribute information is extracted from the initial marine synthetic aperture radar image, including target category, polarization, background brightness, azimuth angle, or time step.

[0112] The attribute information is then normalized, and a neural network embedding layer is used to map the normalized attribute information to a high-dimensional feature space to obtain attribute feature vectors. The mathematical expression for the normalization process is: In the formula, x′ is the normalized value, X is the original value of the corresponding data, and X max This represents the maximum value among all values ​​in the corresponding data. Furthermore, since attributes such as target category, polarization mode, and background brightness are discrete values, neural network embedding layers can be used to map these discrete attributes to a high-dimensional feature space, facilitating model processing and learning. The mathematical expression for target category embedding is: v cat =E cat (A category In the formula, v cat E represents the attribute feature vector of the target category. cat A is the neural network embedding function for the target category. category For the target category. The mathematical expression for polarization embedding is: V pplar =E polar (A polar In the formula, v polarE is the attribute feature vector of the polarization mode. polor For a polarization-based neural network embedding function, A polar This is a polarization method. The mathematical expression for background brightness embedding is: v back =E back (A background In the formula, v back E is the feature vector of the background brightness attribute. back A is the neural network embedding function for background brightness. background For the background brightness, since the azimuth attribute has continuous values ​​distributed in [0°, 360°), sine and cosine encoding is first used, followed by a linear fully connected layer to map it to a high-dimensional feature space. This method can not only represent the magnitude of the azimuth angle but also naturally handle the periodicity of the angle. Its mathematical expression is: v az =E az ([sin(A azimuth ),cos(A azimuth In the formula, v az E is the attribute feature vector of the azimuth angle. az For a fully connected mapping layer in a neural network, A azimuth It is the azimuth angle.

[0113] Finally, multiple attribute feature vectors are concatenated with the first vector sequence to obtain the second vector sequence. The second vector sequence is then embedded with positional encoding to obtain the target feature vector sequence. For example, in order to input the positional information between image patches into the network so that the model can better understand the structure of the image sequence, positional encoding embedding can also be added. The information of each position is encoded into a learnable vector and combined with the 768-dimensional vector sequence after the noisy image patch is transformed to obtain the target feature vector sequence, which can be used as the input of the subsequent network model.

[0114] In some embodiments, in step S103, constructing a set of strong scattering points based on the initial marine synthetic aperture radar image may include, but is not limited to, the following steps:

[0115] Initialize the set of strong scattering points;

[0116] Mean filtering is applied to the initial synthetic aperture radar image at sea to obtain the filtered image;

[0117] Add all pixels in the filtered image to the set of strong scattering points;

[0118] The pixel with the largest value in the filtered image is selected as the strong scattering point;

[0119] A strong scattering region is generated based on the strong scattering point and the preset side length;

[0120] Delete multiple target candidate points from the set of strong scattering points, where each target candidate point corresponds to a pixel in the strong scattering region;

[0121] Return to the step of performing mean filtering on the initial marine synthetic aperture radar image to obtain the filtered image, until the number of strong scattering points in the set of strong scattering points reaches the preset number.

[0122] In some embodiments, considering the interference of speckle noise in the initial marine synthetic aperture radar image, which affects the model's ability to model the characteristics of strong target scattering points, a loss weight matrix designed based on the location of the strong target scattering points can be introduced into the model's loss function calculation process to utilize these characteristics and guide the model to generate more realistic SAR target images. The process for constructing the loss weight matrix based on the strong target scattering points is as follows: Figure 4 As shown. First, a set of strong scattering points can be initialized. Then, the initial marine synthetic aperture radar image is subjected to mean filtering to obtain a filtered image. For example, the initial marine synthetic aperture radar image can be subjected to mean filtering, with the filter size *l* being 1 / 10 of the image size. The value of each pixel in the resulting filtered image represents the average scattering intensity within a *l*×* region centered on that point. Each pixel can be considered a candidate for a strong scattering point. All pixels in the filtered image are then added to the set of strong scattering points, and the pixel with the largest value is selected as the strong scattering point. Based on the strong scattering point and a preset side length *l*, a strong scattering region is generated, with the strong scattering point located at the center of the strong scattering region. Then, multiple target candidate points in the set of strong scattering points are deleted. The target candidate points correspond to pixels in the strong scattering region. For example, the corresponding candidate points in the set of strong scattering points can be deleted based on the pixels in the strong scattering region. The process of performing mean filtering on the initial marine synthetic aperture radar image to obtain the filtered image is repeated until the number of strong scattering points in the set of strong scattering points reaches a preset number. For example, the extraction can be repeated until the number of strong scattering points in the set of strong scattering points reaches 9, thus obtaining the specific locations of the 9 strong scattering points in the original image.

[0123] In some embodiments, in step S104, constructing the loss weight matrix based on the set of strong scattering points may include, but is not limited to, the following steps:

[0124] Determine the location of the strong scattering points based on the set of strong scattering points;

[0125] Distortion detection processing is performed on multiple strong scattering point locations to obtain the noise addition time step size;

[0126] The loss weight matrix is ​​constructed based on the noise addition time step size.

[0127] In some embodiments, the locations of strong scattering points can be determined first based on the set of strong scattering points. For example, the visualization results of strong scattering point extraction from a maritime ship target are shown below. Figure 5 As shown, the red dots represent the specific locations of strong scattering points in the SAR image obtained using a greedy strategy. Then, distortion detection processing is performed on multiple strong scattering point locations to obtain the noise addition time step size. For example, based on the strong scattering point locations, the distortion of scattering points in the image under different noise addition time steps can be analyzed. The distortion detection results show that when the number of noise addition steps reaches 200, the geometric features of the ship target in the image are basically unobservable, submerged in noise. Therefore, a noise addition time step size of 200 steps can be selected, meaning that only the noise image before 200 steps is subjected to loss function weighting processing. Then, based on the noise addition time step size, a loss weight matrix is ​​constructed, where the expression for the loss weight matrix is: w = max(4e -t -3.3,0)+1, where w is the loss weight matrix and t is the noise-adding time step size. More specifically, during the loss function calculation for images with a noise-adding time step size of less than 200 steps, the loss weight matrix w is activated to weight the loss function of the strong scattering points corresponding to the target in the image. Specifically, during the training of the generative diffusion sub-model, the weight matrix w weights the loss function between the model's predicted noise and the actual noise. Compared to non-strong scattering points, due to the effect of the weight matrix, the greater the difference between the predicted noise and the actual noise of strong scattering points, the greater the change in the loss function, causing the model to focus more on extracting the feature information of strong scattering points during the learning process of extracting feature information from the original image.

[0128] In some embodiments, in step S105, constructing the target loss function based on the loss weight matrix may include, but is not limited to, the following steps:

[0129] The target loss function is constructed based on the loss weight matrix, time step, target category attribute condition, azimuth attribute condition, polarization mode attribute condition, and background brightness attribute condition.

[0130] In some embodiments, to ensure the generated marine SAR images possess high quality and accurate multi-attribute combined control, this embodiment incorporates a weighted loss function for combined optimization during model training, combining a strong scattering point weighting scheme. The training process can be simplified as follows: given an input image x0, the image diffusion algorithm progressively adds noise to the image, generating a noisy image x. i , where i represents the number of times noise is added. A set of multi-attribute conditions is also given, including the time step t, the target category attribute condition c. cat Azimuth attribute condition c az Polarization mode attribute condition c polarand background brightness attribute condition c back Image diffusion algorithms need to learn a network ∈ θ To predict what to add to the noisy image x i Noise on the surface. A target loss function can be constructed by weighting the loss weight matrix, time step, target category attribute conditions, azimuth attribute conditions, polarization mode attribute conditions, and background brightness attribute conditions using a strong scattering point guidance scheme. The expression for the target loss function is: In the formula, Let x0 be the target loss function, t be the time step, and c be the input image. cat For the target category attribute condition, c az For the azimuth attribute condition, c polar For polarization mode attribute conditions, c back Given the background brightness attribute, w is the loss weight matrix designed based on strong scattering points. For the noisy image x i The real noise contained within, The model is based on various input attribute conditions and the noisy image x. i The obtained noise prediction results.

[0131] In some embodiments, step S106 includes forward noise addition and denoising, which involves adding and removing noise from the target feature vector sequence. This may include, but is not limited to, the following steps:

[0132] Calculate the variance based on the noise variance coefficient and the identity matrix;

[0133] Calculate the mean based on the noise variance coefficient and the image from the previous step;

[0134] Calculate the conditional probability distribution based on the variance, mean, and the current step image;

[0135] Based on the noise addition time step size and conditional probability distribution, calculate the probability distribution from the original image to the noisy image;

[0136] The noisy image is calculated based on the probability distribution from the original image to the noisy image and the noise following a Gaussian distribution.

[0137] In some embodiments, to improve the quality and diversity of the generated images, this embodiment selects a generation diffusion sub-model as the main structure of the generation process. The noise addition and denoising process includes forward noise addition, which involves adding fixed noise to the image, gradually transforming it into standard Gaussian noise. The noise addition and denoising process is as follows: Figure 6 As shown, from x0 to x TThis is a progressive noise addition process where the noise is known. This process gradually adds noise to the original image, reducing it to pure Gaussian noise. When the noise addition and denoising process is a forward noise addition process, the target feature vector sequence is subjected to forward noise addition, which can be done by first determining the noise variance coefficient β. t Given the identity matrix I, calculate the variance β. t I, and based on the noise variance coefficient β t And the previous image x t-1 Calculate the mean Then, based on the variance, mean, and the current step image, calculate the current image x. t The conditional probability distribution, where the formula for calculating the conditional probability distribution is: In the formula, q(x) t |x t-1 ) represents the current image x t The conditional probability distribution of x t For the current image, x t-1 For the previous image, β t Let I be the noise variance coefficient, and I be the identity matrix. The conditional probability distribution is Gaussian, with a mean of 1 / 2. Then, based on the noise-adding time step size T and the conditional probability distribution, calculate the probability distribution q(x) from the original image to the noisy image. 1:T |x0), representing the noisy image x obtained after T steps of noise addition, starting from the original image x0. T The probability distribution of the image from the original image to the noisy image is calculated using the following formula: In the formula, q(x) 1:T |x0) represents the probability distribution from the original image to the noisy image, and T is the step size of the noise addition time step. Finally, based on the probability distribution from the original image to the noisy image and the Gaussian-distributed noise, the noisy image is calculated. Wherein, β varies with different t values. t The values ​​are predefined and gradually increase from time 1 to T in a linear increment, satisfying: β1 < β2 < ... < β T Simultaneously define: α t =1-β t , The noise addition process must satisfy the specific formula: In the formula, Noise that follows a Gaussian distribution.

[0138] In some embodiments, step S106, the noise addition and denoising process includes inverse denoising processing, which involves adding and denoising noise to the target feature vector sequence, and may include, but is not limited to, the following steps:

[0139] Calculate the probability distribution from the noisy image to the original image;

[0140] The original image is calculated based on the probability distribution from the noisy image to the original image, random Gaussian noise, and weighting coefficients.

[0141] In some embodiments, to improve the quality and diversity of the generated images, this embodiment selects to use a generation diffusion sub-model as the main structure of the generation process. The noise addition and denoising process includes inverse denoising process. The inverse denoising process is to gradually predict and remove noise information from standard Gaussian noise through the model network, and finally generate a high-quality image. Figure 6 In the middle, x T The process up to x0 is a stepwise denoising process, restoring the original image from random noise input. This process requires using a generative diffusion sub-model architecture to predict noise within the noisy image, thereby achieving denoising and ultimately restoring the original image. When the noise addition and denoising are performed inversely, the target feature vector sequence is denoised inversely. First, the probability distribution from the noisy image to the original image is calculated. Then, based on the probability distribution from the noisy image to the original image, random Gaussian noise, and weight coefficients, the original image is calculated. For example, the inverse denoising process mainly attempts to use p(x... t-1 |x t The inverse process of approximating and adding noise, q(x) t-1 |x t Ultimately, this achieves the goal of processing noisy images x. T The ability to gradually restore the original image. This can be obtained from the forward noise addition process: q(x) t-1 |x t ,x0)=N(x t-1 |μ t (x t ,x0),β t I), where q(x) t-1 |x t Let x(x0) be the probability distribution from the noisy image to the original image, where x0 is the original image and x... t For the current image, x t-1 For the previous image, β t Let be the noise variance coefficient, and I be the identity matrix. The inverse denoising update formula is derived as follows: In the formula, ò θ (x t ,t) is a common noise prediction model used in generation-diffusion models, z is random Gaussian noise, and σ t These are adjustable weighting coefficients.

[0142] In some embodiments, in step S106, the marine synthetic aperture radar image generation model can be trained according to the target loss function. During the training process, the model input is a noisy SAR image, and the specific noisy process is as follows: Figure 6As shown in the diffusion-based noise addition method, the output is the predicted noise. The noise signal contained in the noisy image is known noise. Based on the model's predicted noise and the actual noise in the noisy image, a target loss function is calculated and used for backpropagation of the model's gradient during training. The parameters in the model's network architecture are continuously updated until the loss function converges, ultimately resulting in a marine synthetic aperture radar image generation model capable of predicting noise in noisy images.

[0143] In some embodiments, the process of constructing the marine synthetic aperture radar image generation model in step S106 may include, but is not limited to, the following steps:

[0144] Construct a shallow network, which consists of multiple Transformer layers;

[0145] After the shallow network, build the Transformer module;

[0146] After the Transformer module, a deep network is built, which makes skip connections with the shallow network. The deep network consists of multiple Transformer layers.

[0147] After the deep network, a normalization layer is built;

[0148] After the normalization layer, a convolutional layer is constructed.

[0149] In some embodiments, in constructing a marine synthetic aperture radar (SAR) image generation model, a shallow network, a Transformer module, a deep network, a normalization layer, and a convolutional layer can be constructed sequentially. The shallow network includes multiple Transformer layers, the deep network includes multiple Transformer layers, and the deep network and shallow network are connected in skip connections. For example, in the architecture of the marine SAR image generation model, the network architecture composed of multiple Transformer layers can be viewed as a U-shaped network. Figure 2 The example shows a 7-layer Transformer module, but a 13-layer Transformer module can also be used in actual model construction. In the actual model, the first 6 Transformer layers can be called shallow networks, and the last 6 layers can be called deep networks. Taking the last dimension of the matrix as a reference, the outputs of each Transformer layer in the shallow network are aligned and concatenated with the inputs of each Transformer layer in the deep network, serving as the inputs for each Transformer layer in the deep network. This operation of using the output of the shallow layer as part of the input of the deep layer is called a skip connection.

[0150] In some embodiments, in step S107, a target marine synthetic aperture radar image can be generated using a trained marine synthetic aperture radar image generation model. For example, this embodiment employs a multi-attribute condition combination control strategy to stitch together various attribute conditions (target category, azimuth, etc.) as input to a noise prediction network composed of multiple Transformer layers, controlling the network to output noise prediction results. Simultaneously, the noisy image is used as input to the noise prediction network, enabling the network to predict noise output. The noisy input image is updated according to the inverse denoising update formula of the diffusion model. The updated noisy image is then used as input, and the above operation is repeated approximately 50 times to obtain the final target marine synthetic aperture radar image.

[0151] In some embodiments, to verify the data generation effect and reliability of the embodiments of the present invention, training and testing were performed on the measured SAR ship target image dataset—the FUSARShip dataset. After preprocessing the dataset, the final dataset contains 6 ship subclasses of SAR targets, azimuth angles distributed in the range of [0°, 360°), 2 different polarization modes, and 3 different background categories. The flowchart for obtaining the attributes of various SAR targets is shown below. Figure 7 As shown. Among them, (1) the specific process of obtaining the azimuth information of the target image is as follows: First, the rotation detection box of the target is initially obtained by using a rotating target detector. Then, the target orientation angle is calculated based on the rotation detection box as the initial azimuth. Finally, the data of the azimuth information that is not correctly calibrated is corrected by manual correction. (2) The specific process of obtaining the background brightness information is as follows: First, the energy feature statistical calculation and analysis of the image is performed to obtain a set of statistical data. Then, the statistical data is dimensionality reduced to obtain 3 representative features. Finally, the dimensionality reduced features are clustered to divide the image dataset into 3 categories.

[0152] To analyze the impact of strong scattering point feature guidance schemes on the performance of marine SAR image generation results, this embodiment visually compares the image generation results without the feature guidance scheme with those using the scheme. The comparison results of the strong scattering point feature-guided weighted and unweighted schemes are as follows: Figure 8 As shown, compared to the generated image results without using the strong scattering point feature-guided weighting scheme, the generated image using the weighting scheme has better restoration of ship texture details and is closer to the visual features of the target in the original image.

[0153] Furthermore, this embodiment also calculates the multi-scale structural similarity index (MS-SSIM) of the model generation results before and after using the weighted scheme. This index is commonly used for image quality assessment and can be used to measure the similarity between two images. The higher the MS-SSIM value, the higher the similarity between the generated image and the original image. The comparison results of MS-SSIM index under different categories are as follows: Figure 9 As shown. (Summary) Figure 8 and Figure 9 The results show that the model guided by strong scattering point features performs better in modeling target details such as texture. Under each ship target category, the MS-SSIM value of the model incorporating the strong scattering point weighting scheme is higher than that of the model without it, resulting in images that are closer to the original images and possess higher fidelity.

[0154] This embodiment also verifies the model's ability to generate diverse results. The generation results are controlled by combining multiple attributes of SAR images, such as target category, azimuth, polarization, and background brightness. Figure 10 As shown, the generated results include a combination of four attribute information: target category, azimuth, polarization mode, and background category. Specifically, the target category includes six types of SAR ships: Bulk Carrier, Cargo Ship, Container, Fishing, Law Enforcement, and Others. The azimuth includes 12 angles taken at 30° intervals between [0°, 360°). There are two polarization modes (DH and DV) and three background categories (Category I, Category II, and Category III). Observation results show that the generated SAR image has good fidelity and has the following significant advantages: In terms of target feature generation, it can show the texture changes of different ship targets and has various target features and strong scattering point features in the actual test scene; in terms of multi-attribute control, the change in azimuth can be intuitively seen from the direction of the target, and the change in overall image brightness from bright to dark also reflects the change in background. This fully demonstrates that the method proposed in this invention can not only accurately generate the texture and scattering point features of the target, but also maintain the high quality of the generated results and the reliability of the combined generation when multiple attributes are specified at the same time.

[0155] In some embodiments, this embodiment designs a generation scheme guided by the strong scattering point features of SAR targets. Since the strong scattering points of targets in SAR images possess rich target feature information (such as the relative positions of reflective objects within the target), a weighted matrix of the loss function based on the strong scattering points of the targets is designed to enhance the model's ability to model the feature information of the strong scattering points. This embodiment designs an architecture including a multi-attribute combination control module. This architecture can be used to process multiple attributes of marine SAR images and use them as conditional inputs to the model. Based on the characteristics of each attribute, an appropriate embedding encoding method is selected, enabling it to independently and flexibly guide the generation process, ensuring that each attribute is correctly reflected in the generated marine SAR image. This embodiment designs a multi-attribute marine SAR image dataset construction scheme, using various techniques to preprocess the original marine SAR image dataset to obtain more attribute information from the images. For the azimuth information of marine SAR images, the target orientation angle is initially calculated based on the detection box obtained from the rotated target detection model; for the background brightness information of marine SAR images, three different background brightness categories are initially divided based on the energy clustering method.

[0156] In some embodiments, existing marine SAR image generation technologies struggle to balance image generation quality and attribute controllability, failing to fully consider the diverse attribute conditions of marine SAR images. Furthermore, the stability and generalization of model training are weak, neglecting the key generation of strong scattering points from targets. Existing marine SAR image simulation technologies struggle to model image data consistent with real-world application scenarios, rendering them unsuitable for downstream tasks. In contrast, the strong scattering point feature-guided marine SAR image generation method proposed in this embodiment employs a combination of Transformer and diffusion models as the generation network structure, effectively ensuring the quality and diversity of generated marine SAR images. It also introduces a loss weighting matrix designed based on target strong scattering point features to enhance the model's ability to generate and model strong scattering points from targets. Simultaneously, an architecture including a multi-attribute combination control module is designed to achieve controllable generation of multi-attribute combinations in marine SAR images. This allows for the combined processing of attributes such as target category, azimuth, polarization, and background, significantly improving the accuracy of each attribute-guided generation result and the consistency of multi-attribute combination generation, providing new ideas and solutions for research in related fields. Moreover, this model performs excellently on real-world marine SAR datasets, stably generating high-quality images with multi-attribute combinations. These advantages enable the present invention to generate highly usable marine SAR images, effectively alleviating the problem of insufficient measured sample data, and also providing solid technical support and reliability for SAR image generation and its application in downstream tasks such as SAR image interpretation and automatic target recognition.

[0157] The beneficial effects of implementing the embodiments of the present invention include: This application provides a method and apparatus for generating marine synthetic aperture radar images guided by strong scattering point features. The embodiments of this application first acquire an initial marine synthetic aperture radar image, preprocess the initial marine synthetic aperture radar image using a multi-attribute combination control strategy to obtain a target feature vector sequence, then construct a set of strong scattering points and a loss weight matrix based on the initial marine synthetic aperture radar image, and then construct a target loss function based on the loss weight matrix to train the marine synthetic aperture radar image generation model, and finally use the trained marine synthetic aperture radar image generation model to generate the target marine synthetic aperture radar image, thereby enabling the generation of marine synthetic aperture radar images guided by strong scattering point features, improving image quality and reducing computational resource consumption.

[0158] like Figure 11 As shown, this embodiment of the invention also provides a marine synthetic aperture radar image generation device guided by strong scattering point features, comprising:

[0159] The first module 801 is used to acquire initial marine synthetic aperture radar images;

[0160] The second module 802 is used to preprocess the initial marine synthetic aperture radar image using a multi-attribute combined control strategy to obtain a target feature vector sequence.

[0161] The third module 803 is used to construct a set of strong scattering points based on the initial marine synthetic aperture radar image;

[0162] Module 4, 804, is used to construct the loss weight matrix based on the set of strong scattering points.

[0163] Module 5, 805, is used to construct the target loss function based on the loss weight matrix;

[0164] The sixth module 806 is used to train the marine synthetic aperture radar image generation model according to the target loss function. The marine synthetic aperture radar image generation model includes a generation and diffusion sub-model, which is used to add noise and denoise the target feature vector sequence.

[0165] Module 7, 807, is used to generate target marine synthetic aperture radar images using a trained marine synthetic aperture radar image generation model.

[0166] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0167] like Figure 12As shown, embodiments of the present invention also provide a computer device, including:

[0168] At least one processor 901;

[0169] At least one memory 902 is used to store at least one program;

[0170] When at least one program is executed by at least one processor, the at least one processor performs the method described above.

[0171] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0172] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0173] The content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0174] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0175] The content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0176] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0177] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0178] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for generating marine synthetic aperture radar images guided by strong scattering point features, characterized in that, Includes the following steps: Acquire initial synthetic aperture radar images of the sea; The initial marine synthetic aperture radar image is preprocessed using a multi-attribute combined control strategy to obtain a target feature vector sequence; Based on the initial marine synthetic aperture radar image, a set of strong scattering points is constructed; Based on the set of strong scattering points, construct the loss weight matrix; Based on the loss weight matrix, construct the target loss function; Based on the target loss function, a marine synthetic aperture radar image generation model is trained. The marine synthetic aperture radar image generation model includes a generation-diffusion sub-model, which is used to add noise and denoise the target feature vector sequence. The trained marine synthetic aperture radar image generation model is used to generate a target marine synthetic aperture radar image.

2. The method according to claim 1, characterized in that, The preprocessing of the initial marine synthetic aperture radar image using a multi-attribute combined control strategy yields a target feature vector sequence, including: The initial marine synthetic aperture radar image is divided into blocks and encoded to obtain a first vector sequence; Attribute information is extracted from the initial marine synthetic aperture radar image, including target category, polarization mode, background brightness, azimuth angle, or time step. The attribute information is normalized. The normalized attribute information is mapped to a high-dimensional feature space using a neural network embedding layer to obtain the attribute feature vector; The first vector sequence is concatenated with multiple attribute feature vectors to obtain a second vector sequence; The second vector sequence is embedded with position encoding to obtain the target feature vector sequence.

3. The method according to claim 1, characterized in that, The step of constructing a set of strong scattering points based on the initial marine synthetic aperture radar image includes: Initialize the set of strong scattering points; The initial marine synthetic aperture radar image is subjected to mean filtering to obtain a filtered image; Add all pixels in the filtered image to the set of strong scattering points; The pixel with the largest value is selected as the strong scattering point from the filtered image; A strong scattering region is generated based on the strong scattering point and the preset side length; Delete multiple target candidate points from the set of strong scattering points, where the target candidate points correspond to the pixels in the strong scattering region; The process continues until the number of strong scattering points in the set of strong scattering points reaches a preset number.

4. The method according to claim 1, characterized in that, The step of constructing a loss weight matrix based on the set of strong scattering points includes: The location of the strong scattering points is determined based on the set of strong scattering points. Distortion detection processing is performed on the locations of multiple strong scattering points to obtain the noise addition time step size; The loss weight matrix is ​​constructed based on the noise addition time step size.

5. The method according to claim 1, characterized in that, The step of constructing the target loss function based on the loss weight matrix includes: The target loss function is constructed based on the loss weight matrix, time step, target category attribute condition, azimuth attribute condition, polarization mode attribute condition, and background brightness attribute condition.

6. The method according to claim 1, characterized in that, The noise addition and denoising process includes forward noise addition, and the noise addition and denoising process on the target feature vector sequence includes: Calculate the variance based on the noise variance coefficient and the identity matrix; Calculate the mean based on the noise variance coefficient and the image from the previous step; Calculate the conditional probability distribution based on the variance, the mean, and the current step image; Based on the noise addition time step size and the conditional probability distribution, calculate the probability distribution from the original image to the noisy image; The noisy image is calculated based on the probability distribution from the original image to the noisy image and the noise following a Gaussian distribution.

7. The method according to claim 1, characterized in that, The noise addition and denoising process includes inverse denoising processing. The noise addition and denoising process on the target feature vector sequence includes: Calculate the probability distribution from the noisy image to the original image; The original image is calculated based on the probability distribution from the noisy image to the original image, the random Gaussian noise, and the weighting coefficients.

8. The method according to claim 1, characterized in that, The construction process of the marine synthetic aperture radar image generation model includes: Construct a shallow network, which includes multiple Transformer layers; After the shallow network, a Transformer module is built; After the Transformer module, a deep network is constructed, which makes skip connections with the shallow network. The deep network includes multiple Transformer layers. A normalization layer is constructed after the deep network; After the normalization layer, a convolutional layer is constructed.

9. A marine synthetic aperture radar image generation device guided by strong scattering point features, characterized in that, include: The first module is used to acquire initial marine synthetic aperture radar images; The second module is used to preprocess the initial marine synthetic aperture radar image using a multi-attribute combined control strategy to obtain a target feature vector sequence. The third module is used to construct a set of strong scattering points based on the initial marine synthetic aperture radar image; The fourth module is used to construct a loss weight matrix based on the set of strong scattering points; The fifth module is used to construct the target loss function based on the loss weight matrix; The sixth module is used to train the marine synthetic aperture radar image generation model according to the target loss function. The marine synthetic aperture radar image generation model includes a generation-diffusion sub-model, which is used to add noise and denoise the target feature vector sequence. The seventh module is used to generate target marine synthetic aperture radar images using the trained marine synthetic aperture radar image generation model.

10. A computer device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method as described in any one of claims 1-8.

Images