A SAR image self-supervised pre-training method based on a mask autoencoder

Abstract

The present application relates to the technical field of SAR image processing and self-supervised learning. A SAR image self-supervised pre-training method based on a mask autoencoder. In view of the problem that existing SAR target detection relies on natural image pre-training weights and there is a significant domain gap, resulting in low detection accuracy in a small sample scene, a multi-source SAR pre-training dataset is constructed, an asymmetric encoding and decoding architecture is adopted, ResNet50 is used as a feature extraction encoder to adapt to the local scattering characteristics of SAR, a lightweight convolutional decoder is constructed, a Gamma distribution speckle noise modeling mechanism is introduced, the input image is divided into non-overlapping image blocks and a random mask matrix is generated, the encoder only encodes the visible area, the decoder maps the latent representation back to the image space to complete the reconstruction, and the learning rate scheduling and gradient clipping mechanisms are combined to complete unsupervised pre-training, and the encoder weight is extracted and migrated to a downstream rotating target detection model to complete small sample fine tuning.

Description

Technical Field

[0001] This invention relates to the fields of SAR (Synthetic Aperture Radar) image processing, deep learning, self-supervised learning, and remote sensing image interpretation, and provides a SAR image self-supervised pre-training method based on a mask autoencoder. Background Technology

[0002] Synthetic Aperture Radar (SAR), with its all-weather, all-time imaging advantages and ability to penetrate clouds, rain, and snow, has become an indispensable means of Earth observation in the field of remote sensing. Its images can provide high-resolution information on surface targets, providing important support for decision-making in various industries. With the development of deep learning technology, significant progress has been made in SAR image target detection, classification, and other interpretation tasks based on deep learning. However, these methods often rely on a large amount of high-quality labeled data for model training.

[0003] However, SAR image annotation faces challenges such as high costs, high technical barriers, and geographically limited data acquisition, resulting in a scarcity of high-quality annotated data. This is especially true in scenarios with few samples, where model training faces severe data bottlenecks and generalization capabilities decline significantly. To alleviate this problem, the pre-training-fine-tuning paradigm is widely used. This involves pre-training on a large-scale dataset and then transferring the pre-trained weights to downstream SAR tasks for fine-tuning.

[0004] Currently, most mainstream pre-training methods are based on supervised pre-training on natural image datasets such as ImageNet. However, SAR images and natural images have fundamental differences in imaging mechanisms: SAR images have speckle noise, special scattering mechanisms, and unique texture characteristics. The features learned by the pre-trained models of natural images do not match the characteristics of SAR targets, resulting in a significant domain gap. This leads to poor pre-training weight transfer performance and makes it difficult to fully support downstream SAR image interpretation tasks.

[0005] In recent years, the masked image modeling (MIM) paradigm in self-supervised learning, especially the masked autoencoder (MAE), has become a core solution to the data scarcity problem due to its advantages of not requiring manual annotation and being able to learn effective representations from large-scale unlabeled data. MAE forces the model to learn the global semantics and robust features of the image by randomly masking most of the input image and encoding and reconstructing the masked region based only on the visible region. However, existing MAE methods often directly adopt the parameter configuration and training strategies of natural images without fully considering the speckle noise, geometric distortion, and scattering statistics of SAR images. This results in poor adaptability of pre-trained features to downstream SAR tasks, making it difficult to effectively suppress noise interference and capture the scattering structure features of SAR targets.

[0006] Furthermore, existing SAR image pre-training methods either rely on large amounts of labeled data or lack a pre-training architecture specifically designed for SAR characteristics, failing to fully utilize the value of massive amounts of unlabeled SAR data and making it difficult to improve downstream detection accuracy in low-sample scenarios. Therefore, designing a MAE pre-training framework adapted to the characteristics of SAR images, utilizing unlabeled data to learn robust features specific to the SAR domain, bridging the domain gap, and improving the performance of downstream tasks in low-sample scenarios has become a pressing technical challenge. Summary of the Invention

[0007] The purpose of this invention is to solve the domain gap problem caused by the essential differences between the pre-trained weights of natural images and SAR images in imaging mechanisms (speckle noise, scattering mechanism, texture characteristics), which leads to poor adaptability of pre-trained features to downstream SAR tasks and low target detection accuracy in scenarios with few samples.

[0008] To achieve the above objectives, the present invention employs the following technical means:

[0009] This invention provides a self-supervised pre-training method for SAR images based on a mask autoencoder, comprising the following steps:

[0010] Step 1: Construction and preprocessing of multi-source SAR pre-training dataset. Integrate unlabeled SAR data of different bands, resolutions and polarization modes, complete unified format conversion and standardization processing, and construct a large-scale domain-specific pre-training dataset.

[0011] Step 2: SAR-MAE self-supervised pre-training framework construction. An asymmetric encoding and decoding architecture is adopted. ResNet50 is used as the feature extraction encoder to adapt to the local scattering characteristics of SAR. A lightweight convolutional decoder is constructed. A Gamma distribution speckle noise modeling mechanism is introduced to divide the input image into non-overlapping image blocks and generate a random mask matrix. The encoder only encodes the visible area. The decoder maps the latent representation output by the encoder back to the original image space to complete the reconstruction.

[0012] Step 3: Self-supervised pre-training process and optimization strategy. A high-proportion random masking strategy is adopted, and only the pixel reconstruction error of the masked region is calculated. Combined with learning rate scheduling and gradient clipping mechanism, unsupervised pre-training of large-scale SAR data is completed.

[0013] Step 4: Transferring pre-trained weights to downstream detection tasks. After training, discard the decoder part and extract all weight parameters of the ResNet50 encoder as initialization parameters of the backbone network of the downstream rotating target detection model. Use the full convolution feature to adapt the pre-trained input scale to the downstream high-resolution detection input.

[0014] Step 5: Downstream few-sample target detection fine-tuning: Freeze the backbone network parameters, fine-tune only the classification and regression layers of the detection head, use few-sample labeled samples to complete rapid adaptation, and output the target detection results.

[0015] In the above scheme, step 1 includes the following steps:

[0016] Step 1.1: Multi-source SAR data integration. Collect and integrate publicly available SAR datasets, covering X, C, and L bands, with a resolution range of 0.1m to 25m, including single-polarization, dual-polarization, and fully polarization data;

[0017] Step 1.2: Unified data preprocessing. All SAR images are uniformly scaled to 256×256 resolution. The pixel mean and standard deviation of the global dataset are calculated. Pixel-by-pixel normalization is performed on all images. The formula for pixel-by-pixel normalization is:

[0018]

[0019] in These are the original image pixel values. These are the normalized pixel values. The global pixel mean. The standard deviation is denoted as .

[0020] Step 1.3: Data augmentation strategy configuration. During the pre-training stage, a geometric augmentation method of random horizontal flipping, vertical flipping and 90° multiple rotation is adopted, without introducing color transformation to preserve SAR scattering intensity characteristics.

[0021] Step 1.4: Dataset partitioning. The preprocessed dataset is divided into a pre-training set and a validation set in an 8:2 ratio for model training and reconstruction performance evaluation.

[0022] In the above scheme, step 2 includes the following steps:

[0023] Step 2.1: Encoder Structure Design. A standard ResNet50 convolutional residual network is adopted as the feature extraction encoder. Adaptive adjustments are made to the single-channel grayscale characteristics and local scattering characteristics of SAR images, while retaining its complete convolutional layers, batch normalization layers, and residual connection structure. The number of input channels of the encoder is adjusted from the standard 3 channels to 1 channel to match the input requirements of SAR single-channel grayscale images. The encoder finally outputs the feature map of the fourth residual stage, whose output dimension is strictly aligned with the output dimension of the backbone network of the downstream ARS-DETR detection framework, ensuring seamless transfer of pre-trained weights.

[0024] Step 2.2: Asymmetric Lightweight Decoder Design. Addressing the issues of unreasonable resource allocation and low pre-training efficiency caused by the symmetric encoding / decoding architecture of existing masked autoencoders, this invention designs a functionally decoupled asymmetric lightweight convolutional decoder. Its core design idea is: the primary goal of pre-training is to train a high-quality feature extraction encoder; the decoder only undertakes the auxiliary training task of mapping low-dimensional latent representations back to the original image space. After pre-training, the decoder is completely discarded and does not participate in downstream detection inference.

[0025] The decoder input is the feature map output by the encoder, with dimensions (B, 2048, 8, 8) (B is the batch size). The overall architecture adopts a three-stage structure of "projection dimensionality reduction - multi-level upsampling - output alignment". The signal processing flow is as follows:

[0026] Step 2.2.1. Input projection dimensionality reduction layer: The number of channels is compressed from 2048 to 512 through 1×1 convolution, followed by ReLU activation function, and the output size is (B,512,8,8);

[0027] Step 2.2.2. Four-stage progressive upsampling decoding blocks: Four identical upsampling decoding blocks are concatenated. Each decoding block first doubles the feature map size through bilinear interpolation, then performs local feature enhancement through 3×3 convolution, followed by batch normalization layers and ReLU activation functions, while keeping the number of channels constant at 512. After four stages of upsampling, the feature map sizes become 16×16, 32×32, 64×64, and 128×128 respectively.

[0028] Step 2.2.3. Output projection and size alignment layer: The number of channels is compressed to 1 by 1×1 convolution to match the single-channel characteristics of SAR, and then the size is restored to 256×256 by bilinear interpolation. Finally, the reconstructed SAR image is output.

[0029] The decoder has approximately 10.5M parameters, only 1 / 2.4 of the encoder's parameters, achieving optimal resource allocation of "encoder focusing on features, decoder focusing on reconstruction." This asymmetric architecture synergizes with the ResNet50 encoder: compared to the symmetric encoder-decoder architecture, pre-training computational overhead is reduced by approximately 60%; simultaneously, since the decoder can only recover low-frequency global structures, it forces the encoder to learn more robust discriminative features; after pre-training, encoder weights can be directly and seamlessly transferred to downstream detection frameworks without additional parameter transformation.

[0030] Step 2.3: Mask reconstruction mechanism design, the input image is divided into 16×16 non-overlapping image blocks, a binary mask matrix is ​​generated to randomly occlude most of the image area, and the encoder only encodes the visible area;

[0031] Step 2.4: SAR speckle noise modeling. A speckle noise simulation step is introduced into the input preprocessing process of the pre-trained network. By injecting random multiplicative noise that follows the statistical characteristics of Gamma distribution, the inherent speckle characteristics of SAR images are simulated, forcing the model to learn noise-resistant structural features rather than local noise patterns, thereby improving the robustness and generalization ability of the pre-trained features.

[0032] This invention selects a Gamma-distributed random variable with shape parameter k=2 and scale parameter θ=0.5. The mathematical expectation of the distribution of simulated SAR speckle noise. This ensures that the overall average brightness of the image remains unchanged after noise injection, avoiding damage to the original scattering intensity information and variance. This method can accurately fit the noise statistics of multi-source SAR images. Modern deep learning frameworks (PyTorch / TensorFlow) all provide native Gamma distribution sampling operators, whose computational cost is comparable to that of normal distribution sampling, with no significant performance difference. The complete implementation steps are as follows:

[0033] Step 2.4.1. Generate a standard normal random matrix G with the exact same size as the input image, and a Gamma distribution with shape parameter k=2 and scale parameter θ=0.5;

[0034] Step 2.4.2. Set the noise injection stage before the random masking operation. After performing noise injection on the complete input image, generate a binary mask matrix to randomly occlude most of the image area (the mask area is filled with the global pixel mean).

[0035] Step 2.4.3. According to the formula Perform pixel-wise multiplicative noise injection, where I is the initial image. The image after noise injection;

[0036] Step 2.4.4. Input the image with injected noise into the encoder for feature extraction.

[0037] Noise intensity is positively correlated with image pixel brightness, consistent with the physical imaging mechanism of SAR speckle noise, where "strong scattering points correspond to strong noise." Furthermore, noise interference is introduced only in the masked region, avoiding impact on feature extraction quality in the visible area and forcing the model to ignore high-frequency noise during reconstruction, focusing instead on restoring the target's low-frequency global geometry and scattering features. By simulating speckle noise of varying intensities, the cross-domain generalization ability of pre-trained weights on real SAR images under different sensors and imaging conditions is significantly improved.

[0038] In the above scheme, step 3 includes the following sub-steps:

[0039] Step 3.1: Mask ratio optimization setting, using a mask ratio of 0.6 to balance noise suppression and target structure preservation. This parameter is optimized based on the speckle noise characteristics of SAR images. The optimal value was obtained by comparing pre-training experiments with five different mask ratios (0.5, 0.6, 0.7, 0.75, and 0.8) and verifying the performance of downstream few-sample detection tasks. Compared to the high mask ratios of 0.75~0.9 used for natural images (low noise, high effective information density), SAR images have significantly reduced effective information density due to inherent speckle noise, resulting in a much greater amount of missing information at the same mask ratio compared to natural images. 0.6 is the optimal high-ratio mask value for SAR images: this ratio ensures that the model must rely on global structural features for reconstruction, avoiding fitting local noise patterns, while also preventing the model from failing to converge due to excessive information loss. The masking rate of 0.6 and the Gamma distribution speckle noise modeling in step 2.4 (adding noise first and then masking) have a synergistic effect, jointly forcing the model to learn noise-resistant global scattering structure features.

[0040] Step 3.2: Loss function definition, using mean squared error as the reconstruction loss, only calculating the pixel reconstruction error of the masked region, without constraining the visible region; Mean squared error loss for the masked region:

[0041]

[0042] in The total number of pixels being masked. The set of pixel indices for the masked region. The pixel values ​​reconstructed by the model. These are the actual pixel values ​​of the original image. Only the loss in the masked region is calculated; the visible region is not constrained.

[0043] Step 3.3: Optimizer and Learning Rate Scheduling. The AdamW optimizer is used, with an initial learning rate of 1.5 × 10⁻⁴. A linear warm-up mechanism is introduced for the first 2000 iterations, followed by a cosine annealing learning rate scheduling strategy. Multiple comparative experiments were conducted within a reasonable range of 1 × 10⁻² to 5 × 10⁻³ to verify the optimal value obtained for SAR self-supervised pre-training. This value ensures model convergence speed while avoiding feature drift due to an excessively high learning rate or slow convergence due to an excessively low learning rate. The cosine annealing learning rate formula is:

[0044]

[0045] in Let be the learning rate in round t. The initial learning rate, This refers to the total number of training rounds.

[0046] Step 3.4: Training stability control, set the gradient clipping threshold to 1.0 to suppress the gradient explosion problem caused by strong scattering points in SAR images, set the batch size to 40, and the total number of training rounds to 200.

[0047] In the above scheme, step 4 includes the following sub-steps:

[0048] Step 4.1: Pre-trained weight extraction. After training, discard the decoder part used only for the reconstruction task and extract all weight parameters of the ResNet50 encoder.

[0049] Step 4.2: Detection model structure adaptation. The extracted encoder weights are fully loaded into the backbone network of the downstream ARS-DETR detection framework. The neck network, Transformer encoder / decoder, and detection head of the detection framework are all randomly initialized using the Xavier method. When loading pre-trained weights, non-existent keys are ignored by using the strict=False parameter to ensure that only the backbone network loads SAR-MAE pre-trained weights.

[0050] Step 4.3: Input scale adaptation. Utilizing the fully convolutional characteristics of ResNet50, the 256×256 input scale during the pre-training stage is directly adapted to the 1024×1024 high-resolution input of the downstream detection task without adjusting the convolution kernel parameters.

[0051] In the above scheme, step 5 includes the following sub-steps:

[0052] Step 5.1: Full-supervised training of base classes. The transferred detection model is trained in a fully supervised manner using the 17 base class labeled data of the FAIR-CSAR dataset to optimize the joint parameters of the backbone network and the detection head.

[0053] Step 5.2: Few-shot fine-tuning for new classes. Freeze all parameters of the backbone network, Transformer encoder, and neck network, and fine-tune only the classification and regression layers of the detection head. Use 5-shot, 10-shot, or 20-shot of new class labeled samples to complete rapid adaptation; the learning rate for the few-shot fine-tuning stage is set to 1×10⁻ 5 The weight decay is set to 5×10⁻ 4 The number of rounds for fine-tuning is 200.

[0054] Step 5.3: Model Inference. Input the SAR image to be detected, and the model outputs the five parameters of the rotated bounding box of the target. The system outputs the final detection results after filtering out predicted bounding boxes with a confidence score below 0.05, including category labels and confidence scores.

[0055] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:

[0056] 1. This invention addresses the problems of poor model generalization and weak cross-sensor adaptability caused by the single nature of existing SAR self-supervised pre-training datasets, limited scene coverage, and single imaging conditions, by constructing a large-scale multi-source SAR pre-training dataset containing 13 data sources (see step 1). It achieves comprehensive coverage of SAR data in multiple bands such as X, C, and L, multiple resolutions from 0.1m to 25m, and single / dual / full polarization imaging modes. This enables the model to learn general global scattering structure features across scenes and imaging conditions, significantly enhancing the ability to identify targets in different SAR sensors and different geographical environments, and is suitable for complex and ever-changing remote sensing monitoring tasks.

[0057] 2. This invention adopts an asymmetric encoding and decoding architecture based on ResNet50 (see steps 2.1-2.2), which solves the problems of the original MAE framework using Transformer encoders being unsuitable for the local scattering characteristics of SAR images, having high computational complexity, and being difficult to handle large-scale SAR data. It fully leverages the natural advantages of convolutional neural networks in local pattern modeling and translation invariance. At the same time, through the lightweight decoder design, the main computational resources are concentrated on the extraction of high-quality encoded representations, which greatly reduces the computational overhead of pre-training and improves the feasibility and efficiency of pre-training large-scale SAR data.

[0058] 3. This invention designs a collaborative strategy for high-ratio mask reconstruction tasks adapted to SAR imaging characteristics and speckle noise modeling with noise addition followed by masking (see steps 2.3-2.4 and 3.1). This solves the core problems of a significant domain gap between natural image pre-trained weights and the SAR domain, and the model's tendency to learn local speckle noise rather than the true structure of the target. By using an optimal mask ratio of 0.6, the model is forced to ignore high-frequency noise and focus on restoring the low-frequency global geometric structure. At the same time, multiplicative noise with Gamma distribution is introduced to simulate the inherent speckle characteristics of SAR, enabling the model to learn domain-specific scattering representations with strong noise resistance and high discriminativeness. This fundamentally alleviates the feature initialization bottleneck of downstream detection tasks. Furthermore, the use of the physical real operation order of "noise addition followed by masking" avoids the model learning false noise distribution patterns, further improving the cross-domain generalization ability of pre-trained weights on real SAR images under different sensors and imaging conditions.

[0059] 4. This invention employs a targeted SAR pre-training optimization strategy (see steps 3.3-3.4) to address the problems of gradient explosion caused by strong scattering points in SAR images, unstable training process, and slow convergence speed. By introducing a linear warm-up mechanism for the first 2000 iterations, cosine annealing learning rate scheduling, and gradient clipping with a threshold of 1.0, the optimization instability caused by strong scattering points is effectively suppressed, ensuring smooth convergence of the large-scale pre-training process, while improving the quality and generalization ability of the final pre-trained weights.

[0060] 5. The backbone network-only migration strategy proposed in this invention (see step 4) solves the problems of incompatibility between pre-trained weights and downstream rotating object detection frameworks, high migration costs, and input scale mismatch. By retaining the encoder output with the same dimension as the downstream detection framework, it achieves seamless migration of pre-trained weights to mainstream rotating detection frameworks such as ARS-DETR and GFE-DETR. At the same time, by utilizing the fully convolutional characteristics of ResNet50, the 256×256 pre-training scale is directly adapted to the 1024×1024 high-resolution detection input without any parameter adjustment, which greatly reduces the complexity of engineering deployment.

[0061] 6. This invention significantly improves the performance of SAR target detection under few-shot conditions (see step 5), and solves the problems of poor feature initialization quality, easy model overfitting, and difficulty in identifying special structural targets in few-shot scenarios. Experiments show that the SAR-MAE pre-trained weights achieve a relative performance improvement of 69.2% compared with ImageNet pre-training in 20-shot few-shot tasks. In particular, it has made breakthrough progress in SAR special structural targets such as oil tanks and helicopters, which are difficult to represent by natural image pre-training, enabling the model to achieve high-precision target detection with only a small number of labeled samples.

[0062] 7. This invention makes full use of massive unlabeled SAR data for self-supervised pre-training (see steps 1-3), which solves the industry pain points of high cost of SAR image annotation, scarcity of labeled data, and long annotation cycle. It can learn high-quality visual representations without any manual annotation, greatly reducing the dependence on labeled data, effectively alleviating the problem of insufficient labeled resources in the field of remote sensing, and shortening the model development cycle, which is suitable for the practical application needs of rapid iteration.

[0063] 8. This invention has good modular compatibility and scalability (see steps 4-5), and solves the problem of incompatibility between existing pre-training methods and subsequent technology stacks such as fine-grained detection and incremental learning. Its output backbone network features can be directly and seamlessly combined with geometric feature enhancement module and incremental feature distillation module to form a complete technical system from pre-training to few-sample detection and then to incremental learning, which further improves the overall performance of SAR image intelligent interpretation system.

[0064] 9. This invention is specifically designed based on the physical imaging mechanism of SAR images (see steps 2.4 and 3.1), which solves the problem of poor performance when general computer vision methods are directly transferred to the SAR field. From feature extraction and task design to optimization strategy, the entire process is adapted to the scattering imaging characteristics and noise characteristics of SAR, so that the model can truly understand the physical nature of SAR images, rather than simply imitating the processing of natural images. This significantly improves the robustness and reliability of SAR target detection systems and can be widely used in many fields such as military reconnaissance, marine monitoring, disaster assessment, and urban planning.

[0065] In summary, this invention systematically solves the core domain gap problem of mismatch between natural image pre-training and SAR imaging mechanism by constructing a self-supervised pre-training framework for the SAR domain. It fully taps the value of massive unlabeled SAR data, provides a high-quality feature initialization scheme for downstream few-sample target detection tasks, and significantly improves the accuracy and efficiency of intelligent interpretation of SAR images. It has important academic value and broad industrial application prospects. Attached Figure Description

[0066] Figure 1 This is the overall system flowchart of the present invention;

[0067] Figure 2 The diagram below illustrates the proposed SAR-MAE self-supervised pre-training framework and its downstream rotating target detection transfer experiment architecture. The framework is divided into two core modules: the upper module (SAR-MAE self-supervised pre-training) randomly masks the input SAR image, extracts visible block features using a ResNet50 encoder, reconstructs the image by combining a learnable mask token with a lightweight decoder, optimizes the loss through the masked region MSE, and ultimately retains the encoder weights. The lower module (downstream transfer experiment) uses the pre-trained ResNet50 weights as the backbone, loads them into the ARS-DETR detection framework, and sequentially performs base class training and new class few-sample fine-tuning to adapt the pre-trained features to the SAR target detection task. Detailed Implementation

[0068] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0069] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0070] Current SAR target detection heavily relies on high-quality labeled data, while mainstream pre-training schemes directly transfer weights from natural images such as ImageNet. Due to the fundamental differences in imaging mechanisms between SAR images and natural images—SAR images possess speckle noise, unique scattering mechanisms, and distinctive textures—the pre-trained features from natural images do not match the characteristics of SAR targets, resulting in a significant domain gap. This domain gap leads to poor transfer performance of pre-trained weights, resulting in low detection accuracy and a tendency to overfit in scenarios with few samples. Furthermore, existing self-supervised masked autoencoder (MAE) methods directly adopt the parameter configuration and training strategies of natural images, failing to adapt to the statistical characteristics of speckle noise and local scattering patterns in SAR. This makes it difficult to effectively utilize massive amounts of unlabeled SAR data and provide high-quality feature initialization under limited sample conditions.

[0071] To facilitate understanding of the technical concept of this invention by those skilled in the art, the core innovative points are described as follows:

[0072] 1. Asymmetric convolutional encoding / decoding architecture replaces Transformer MAE

[0073] Core technology: Abandoning the computationally expensive Transformer encoder in the original MAE, ResNet50 is used as the feature extraction backbone, along with a lightweight convolutional decoder with only 1 / 10 the number of parameters of the encoder.

[0074] Synergistic effect logic: The local receptive field and translation invariance of ResNet50's convolutional kernels naturally adapt to the local scattering clusters and edge structures of SAR targets, while the lightweight decoder concentrates computational resources on high-quality encoded representation extraction. The two work together to solve the problem of high computational complexity and difficulty in adapting to the local scattering characteristics of SAR by Transformer MAE, and significantly reduce pre-training overhead.

[0075] 2. Cooperative Mechanism between Gamma Distribution Speckle Noise Modeling and High-Scale Mask Reconstruction Adapted to SAR Characteristics

[0076] Core technology: Introducing random multiplicative noise with a Gamma distribution that follows a shape parameter k=2 and a scale parameter θ=0.5 into the input image, first simulating real speckle noise on the complete image, and then combining a random mask ratio of 0.6 with a mechanism that only calculates the MSE loss of the masked region.

[0077] Synergistic Effect Logic: This combination produces a clear, unintended technical effect. Gamma noise simulates the high-frequency interference of inherent speckle noise in SAR, while a high mask ratio of 0.6 forces the model to lack a significant amount of context. The synergy of these two factors compels the model to abandon fitting local speckle noise patterns and instead learn to recover low-frequency global geometry and scattering features. This is a core approach to bridging the gap between natural image pre-training and the SAR domain. Conventional MAE, using a mask ratio of 0.75–0.9 and modeling without physical noise, cannot achieve this noise-resistant representation learning.

[0078] 3. SAR-domain specific training stability control and lossless transfer paradigm at the fully convolutional scale

[0079] Core technologies: A gradient pruning threshold of 1.0 is set for strong scattering points, combined with 2000 iterations of linear warm-up and cosine annealing scheduling; the decoder is discarded after training, and the 256×256 input is directly adapted to the downstream 1024×1024 detection scale by utilizing the fully convolutional characteristics of ResNet50; the backbone network and neck / Transformer encoder are frozen in the few-shot stage, and only the detection head classification and regression layers are fine-tuned.

[0080] Synergistic effect logic: Gradient clipping and learning rate scheduling are deeply coupled with the SAR-MAE reconstruction task, suppressing gradient explosion caused by strong scattering points and ensuring smooth convergence of large-scale pre-training. Fully convolutional scale adaptation eliminates parameter transfer barriers between different resolutions. The few-sample fine-tuning strategy of freezing the backbone and only tuning the detector head, supported by high-quality pre-trained representations, significantly reduces the risk of few-sample overfitting and achieves high-precision and rapid adaptation under conditions of extremely scarce labeled data.

[0081] Furthermore, to facilitate a better understanding of the technical concept of the present invention by those skilled in the art, the technical solution of the present invention is described in detail below:

[0082] This invention addresses the limitation of transfer learning effectiveness caused by significant domain differences between SAR images and natural images, proposing a self-supervised pre-training framework (SAR-MAE) for SAR scenes. This method constructs a large-scale multi-source SAR pre-training dataset containing 13 data sources. Based on the idea of ​​a mask autoencoder, it designs a high-proportion mask reconstruction task and a speckle noise modeling strategy, enabling the model to learn robust global scattering structure features under unlabeled conditions, providing high-quality feature initialization for downstream few-sample target detection tasks.

[0083] This invention is based on the mmrotate framework. The SAR-MAE pre-training module, data preprocessing module, and weight transfer module are all modular structures that can be independently configured and scheduled, supporting seamless integration with existing rotating target detection frameworks. All modules are parameterized through configuration files, allowing adaptation to different scenarios without modifying the core framework code.

[0084] Step 1: Construction and preprocessing of multi-source SAR pre-training dataset. Integrate unlabeled SAR data of different bands, resolutions and polarization modes, complete unified format conversion and standardization processing, and construct a large-scale domain-specific pre-training dataset.

[0085] Step 1.1: Multi-source SAR data integration. Thirteen publicly available SAR datasets were collected and integrated: SAR-AirCraft, SARDet-100k, SSDD, HRSID, MSAR-1.0, Official-SSDD-, OGSOD-1.0, OpenSARShip, OpenSARUrban, RSDD-SAR, SAMPLE-Dataset-Master, Ship-Dataset-v0, and SIVED. These datasets comprehensively cover the X, C, and L bands, with resolutions ranging from 0.1m to 25m. They include single-polarization, dual-polarization, and fully polarized data, covering typical SAR target scenes such as aircraft, ships, vehicles, urban buildings, and ports. Invalid images, including those that are entirely black, entirely white, or have a pixel variance less than 0.01, were removed. Finally, a large-scale unlabeled pre-trained dataset containing 314,254 images was constructed.

[0086] Step 1.2: Unified data preprocessing, scaling all SAR images to a resolution of 256×256, calculating the pixel mean and standard deviation of the global dataset, and performing pixel-by-pixel normalization on all images;

[0087] The specific steps are as follows:

[0088] Step 1.2.1. First, perform the Resize operation, using bilinear interpolation to uniformly scale all images to 256×256 pixels;

[0089] Step 1.2.2. Next, calculate the global pixel mean of the entire pre-training dataset. with standard deviation Pixel-wise normalization is performed through the Normalize operation, using the following formula:

[0090]

[0091] in These are the original image pixel values. These are the normalized pixel values;

[0092] Step 1.3. Data augmentation strategy configuration: During the pre-training phase, only geometric augmentation is used, specifically including: random horizontal flip (execution probability 0.5), random vertical flip (execution probability 0.5), and random rotation by multiples of 90° (execution probability 0.5); no color transformation operations are introduced to avoid damaging the physical characteristics of the scattering intensity of the SAR image.

[0093] Step 1.4. Dataset Splitting and Loading Configuration: The preprocessed dataset is randomly divided into a pre-training set (251,403 images) and a validation set (62,851 images) in an 8:2 ratio. The following settings are configured in the training configuration file:

[0094] Step 1.4.1. The training set batch_size=40, and is loaded using distributed data parallelism (DDP), with 8 samples allocated to each GPU;

[0095] Step 1.4.2. Set the validation set batch_size=80, perform validation every 10 rounds, and calculate the average reconstruction loss;

[0096] Step 1.4.3. Enable pin_memory=True and num_workers=8 to improve data loading speed and avoid GPU waiting.

[0097] Step 2: SAR-MAE self-supervised pre-training framework construction. An asymmetric encoding and decoding architecture is adopted. ResNet50 is used as the feature extraction encoder to adapt to the local scattering characteristics of SAR. A lightweight convolutional decoder is constructed. A Gamma distribution speckle noise modeling mechanism is introduced to divide the input image into non-overlapping image blocks and generate a random mask matrix. The encoder only encodes the visible area. The decoder maps the latent representation output by the encoder back to the original image space to complete the reconstruction.

[0098] A SAR-MAE self-supervised pre-trained network module is constructed, which consists of three parts: an encoder submodule, a lightweight convolutional decoder submodule, and a mask generation submodule. This module fully implements the aforementioned asymmetric encoding and decoding architecture, as detailed below:

[0099] Step 2.1. Encoder Design: A standard ResNet50 convolutional residual network is adopted as the feature extraction encoder, retaining its complete convolutional layers, batch normalization layers, and residual connection structure. Adaptation adjustments are made to the single-channel grayscale characteristics of SAR images, reducing the number of input channels from the standard 3 channels to 1 channel to match the input requirements of SAR single-channel grayscale images. The encoder ultimately outputs the feature map of the fourth residual stage, whose output dimension is strictly aligned with the output dimension of the backbone network of the downstream ARS-DETR detection framework, ensuring seamless transfer of pre-trained weights.

[0100] Step 2.2. Implementation of an Asymmetric Lightweight Convolutional Decoder

[0101] A convolutional decoder with a parameter size significantly smaller than the encoder is constructed, with an optimal parameter ratio of approximately 1:10 to 1:15. The decoder contains only deconvolutional upsampling and normalized activation layers, responsible for progressively restoring low-dimensional latent representations to the original pixel space. After pre-training, this decoder is discarded and does not participate in downstream inference, thus concentrating computational resources on extracting high-quality encoded features. The decoder adopts a hierarchical structure of "input projection - four-level upsampling - output alignment," and the complete implementation logic is as follows:

[0102] Input layer: Receives the 8×8×512 feature map output from the encoder;

[0103] Upsampling layer 1: 2×2 deconvolution, stride 2, output channels 256, followed by BatchNorm2d and ReLU;

[0104] Upsampling layer 2: 2×2 deconvolution, stride 2, output channels 128;

[0105] Upsampling layer 3: 2×2 deconvolution, stride 2, output channels 64;

[0106] Upsampling layer 4: 2×2 deconvolution, stride 2, output channels 32;

[0107] Output layer: 1×1 convolution, output channel 1, generating a 256×256 reconstructed image.

[0108] The decoder is only responsible for the reconstruction task and is discarded directly after pre-training, without participating in downstream detection tasks. Step 2.3. Mask reconstruction mechanism design: The input image is divided into 16×16 non-overlapping image blocks, and a binary mask matrix is ​​generated to randomly occlude most of the image region. The encoder only encodes the visible region, performing random mask generation and visible region encoding according to a mask ratio of 0.6. The specific implementation steps are as follows:

[0109] Step 2.3.1 Divide the input 256×256 SAR image into 16×16 non-overlapping image blocks, resulting in a total of 256 image blocks;

[0110] Step 2.3.2 Generate a random index sequence of length 256, take the image blocks corresponding to the first 154 indices as the mask region, and the remaining 102 as the visible region;

[0111] Step 2.3.3 Generate a binary mask matrix with the same size as the input image, with the masked area marked as 0 and the visible area marked as 1;

[0112] Step 2.3.4 The encoder only extracts features from the pixels corresponding to the visible area; the masked area does not participate in the encoding process.

[0113] Step 2.4: SAR speckle noise modeling. Random multiplicative noise following a Gamma distribution is introduced into the input image to simulate the inherent speckle characteristics of SAR images, forcing the model to learn noise-resistant structural features rather than noise patterns. The Gamma-distributed speckle noise injection is performed according to the following steps:

[0114] Step 2.4.1 sets the noise injection stage after the random masking operation and before the encoder input, performing noise injection only on the masked images; before the forward propagation of each training batch, noise injection is performed on the current batch images with a probability of 0.5. This forces the model to learn noise-resistant structural features rather than local noise patterns. A random variable G is generated following a Gamma distribution with shape parameter k=2 and scale parameter θ=0.5.

[0115] Step 2.4.2 Perform multiplicative noise injection on the input image, using the following formula:

[0116]

[0117] in The original input image, Let be a random variable following a Gamma distribution with shape parameter k=2 and scale parameter θ=0.5. This is the image after noise has been injected.

[0118] Step 3: Self-supervised pre-training process and optimization strategy. A high-proportion random masking strategy is adopted, and only the pixel reconstruction error of the masked region is calculated. Combined with learning rate scheduling and gradient clipping mechanism, unsupervised pre-training of large-scale SAR data is completed.

[0119] Step 3.1: Mask ratio optimization setting, using a mask ratio of 0.6 to balance noise suppression and target structure preservation;

[0120] Step 3.2: Loss function definition. Mean squared error is used as the reconstruction loss. Only the pixel reconstruction error of the masked region is calculated, without constraining the visible region. The formula for the mean squared error loss of the masked region is:

[0121]

[0122] in The total number of pixels being masked. The set of pixel indices for the masked region. The pixel values ​​reconstructed by the model. These are the actual pixel values ​​of the original image. Only the loss in the masked region is calculated; the visible region is not constrained.

[0123] Step 3.3: Optimizer and learning rate scheduling, using the AdamW optimizer, with an initial learning rate set to 1.5 × 10⁻⁶. -4 A linear warm-up mechanism is introduced for the first 2000 iterations, followed by a cosine annealing learning rate scheduling strategy. Specifically:

[0124] Optimizer and learning rate scheduling configuration: Set the optimizer and learning rate strategies in the training configuration file.

[0125] Step 3.3.1: Use the AdamW optimizer with a weight decay coefficient of 0.05 and momentum parameters β1=0.9 and β2=0.999;

[0126] Step 3.3.2: Set the initial learning rate to 1.5 × 10⁻ 4 A linear warm-up mechanism for the first 2000 iterations is introduced, which is implemented through WarmupLRHook;

[0127] Step 3.3.3: After warm-up, cosine annealing learning rate scheduling is used, with a total of 200 training rounds, implemented through CosineAnnealingLRHook. The formula for the learning rate change with the number of training rounds is:

[0128]

[0129] in Let be the learning rate in round t. The initial learning rate, This refers to the total number of training rounds.

[0130] Step 3.4: Training stability control. Set the gradient clipping threshold to 1.0 to suppress gradient explosion caused by strong scattering points in SAR images. Set the batch size to 40 and the total training epochs to 200. Simultaneously configure training process monitoring and optimization strategies to ensure smooth convergence during large-scale pre-training, as follows:

[0131] Step 3.4.1: Set the gradient clipping threshold to 1.0 in the training optimization configuration to limit the gradient norm of all parameters during backpropagation, effectively suppress the gradient explosion problem caused by strong scattering points in SAR images, and ensure the numerical stability of the training process.

[0132] Step 3.4.2: Configure a training loss monitoring mechanism to record and output the average reconstruction loss of the training set and the average reconstruction loss of the validation set every 100 iterations, in order to track the model training status in real time;

[0133] Step 3.4.3: Configure a model weight saving mechanism to save the complete model weights every 10 training rounds; at the same time, automatically record the training round with the lowest reconstruction loss on the validation set and save the optimal model weights for that round separately.

[0134] Step 3.4.4: Configure an early training stop mechanism. When the average reconstruction loss on the validation set no longer decreases after 20 consecutive training rounds, the training process will be automatically terminated to avoid model overfitting and save computational resources.

[0135] Step 4: Transferring pre-trained weights to downstream detection tasks. After training, discard the decoder part and extract all weight parameters of the ResNet50 encoder as initialization parameters of the backbone network of the downstream rotating target detection model. Use the full convolution feature to adapt the pre-trained input scale to the downstream high-resolution detection input.

[0136] Step 4.1: Pre-training weight extraction. After training, discard the decoder part used only for the reconstruction task, extract all weight parameters of the ResNet50 encoder, and complete the extraction and format standardization of the pre-training weights so that the encoder weights can be directly loaded into the backbone network of the downstream detection model. The specific steps are as follows:

[0137] Step 4.1.1: Load the overall weights of the trained SAR-MAE model and extract all parameters of the encoder part; Step 4.1.2: Standardize and adjust the parameter naming of the encoder weights to be consistent with the parameter naming rules of the ResNet backbone network in the downstream detection framework.

[0138] Step 4.1.3: Retain the mean, variance, scaling factor, and offset of the batch normalization layer to ensure the consistency of feature distribution;

[0139] Step 4.1.4: Save the standardized encoder weights as a separate pre-trained weight file for use in the initialization of the downstream detection model.

[0140] Step 4.2: Detection model structure adaptation. The extracted encoder weights are fully loaded into the backbone network of the downstream ARS-DETR detection framework. The neck network, Transformer encoder / decoder, and detection head of the detection framework are all randomly initialized using the Xavier method. The specific implementation is as follows:

[0141] Step 4.2.1: Use the encoder weights obtained from SAR-MAE pre-training as the initial weights of the backbone network of the downstream detection model;

[0142] Step 4.2.2: The neck network, Transformer encoder, Transformer decoder, and detection head of the detection model are all initialized using Xavier uniform distribution.

[0143] Step 4.2.3: When loading weights, only the parameters of the backbone network are matched, and the parameters of the newly added structure of the detection framework are ignored to ensure that only the backbone network loads the SAR-MAE pre-trained weights.

[0144] Step 4.3: Input scale adaptation. Utilizing the fully convolutional characteristics of ResNet50, the 256×256 input scale from the pre-training stage is directly adapted to the 1024×1024 high-resolution input of the downstream detection task without adjusting the convolution kernel parameters. Since the convolution kernel parameters are independent of the input size, no parameter adjustments are needed to the backbone network, thus enabling feature extraction from images of different resolutions.

[0145] Step 5: Downstream few-shot target detection fine-tuning. Freeze the backbone network parameters and fine-tune only the classification and regression layers of the detection head. Use a few labeled samples to complete rapid adaptation and output the target detection results. The specific training configuration is as follows:

[0146] Step 5.1: Full-supervised training of base classes. The transferred detection model is trained in a fully supervised manner using the 17 base class labeled data of the FAIR-CSAR dataset to optimize the joint parameters of the backbone network and the detection head.

[0147] Specifically, the configuration includes a fully supervised training workflow for the base classes, using a subset of the SL dataset from the FAIR-CSAR dataset as the downstream detection task dataset, containing 17 base classes and 5 new classes.

[0148] Step 5.1.1: Load the ARS-DETR model with transferred SAR-MAE pre-trained weights;

[0149] Step 5.1.2: Optimizer configuration: AdamW, initial learning rate 1×10⁻ 4 Weight decay of 0.05, batch size 2;

[0150] Step 5.1.3: The training consists of 36 rounds, using cosine annealing learning rate scheduling;

[0151] Step 5.1.4: The loss function consists of classification cross-entropy loss, rotation GIoU loss and angle classification loss, with weights of 1.0, 2.0 and 0.5, respectively.

[0152] Step 5.2: Fine-tuning with a small number of new class samples. Freeze all parameters of the backbone network, Transformer encoder and neck network, and fine-tune only the classification layer and regression layer of the detection head. Use 5-shot, 10-shot or 20-shot of new class labeled samples to complete the rapid adaptation.

[0153] Specifically, after the base class training is completed, the new class few-shot fine-tuning phase begins with parameter freezing and gradient control.

[0154] Step 5.2.1: After the base class training is completed, the parameters of the backbone network, Transformer encoder and neck network are frozen so that their parameters are no longer updated during the fine-tuning phase;

[0155] Step 5.2.2: Fine-tune only the classification and regression layers of the detection head, setting the learning rate to 1×10⁻ 5 Weight decay 5×10⁻ 4 ;

[0156] Step 5.2.3: Construct fine-tuning datasets with settings of 5-shot, 10-shot, and 20-shot respectively, providing a corresponding number of labeled samples for each class;

[0157] Step 5.2.4: Fine-tune the training rounds to 200 rounds, using cosine annealing learning rate scheduling. During training, only new class samples are used, and base class data is not accessed.

[0158] Step 5.3: Model Inference. Input the SAR image to be detected, and the model outputs the five parameters of the rotated bounding box of the target. The system outputs the final detection results after filtering out predicted bounding boxes with a confidence score below 0.05, including category labels and confidence scores.

[0159] Specifically, the inference process for rotating target detection

[0160] Step 5.3.1: Input the 1024×1024 SAR image to be detected and perform normalization preprocessing consistent with the training phase;

[0161] Step 5.3.2: The model forward propagation outputs 300 target prediction results, including the five parameters of the rotated bounding box. Category labels and confidence scores;

[0162] Step 5.3.3: Filter out prediction boxes with a confidence level below 0.05; no non-maximum suppression (NMS) is required.

[0163] Step 5.3.4: Output the final detection results, which can be saved as a DOTA format annotation file or a visualization image.

[0164] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the concept of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A self-supervised pre-training method for SAR images based on a mask autoencoder, characterized in that, It includes the following steps: Step 1: Construction and preprocessing of multi-source SAR pre-training dataset. Integrate unlabeled SAR data of different bands, resolutions and polarization modes, complete unified format conversion and standardization processing, and construct a large-scale domain-specific pre-training dataset. Step 2: SAR-MAE self-supervised pre-training framework construction. An asymmetric encoding and decoding architecture is adopted. ResNet50 is used as the feature extraction encoder to adapt to the local scattering characteristics of SAR. A lightweight convolutional decoder is constructed. A Gamma distribution speckle noise modeling mechanism is introduced to divide the input image into non-overlapping image blocks and generate a random mask matrix. The encoder only encodes the visible area. The decoder maps the latent representation output by the encoder back to the original image space to complete the reconstruction. Step 3: Self-supervised pre-training process and optimization strategy. A high-ratio random masking strategy adapted to the characteristics of SAR images is adopted. Only the pixel reconstruction error of the masked area is calculated. Combined with learning rate scheduling and gradient clipping mechanism, unsupervised pre-training of large-scale SAR data is completed. Step 4: Transferring pre-trained weights to downstream detection tasks. After training, discard the decoder part and extract all weight parameters of the ResNet50 encoder as initialization parameters of the backbone network of the downstream rotating target detection model. Use the full convolution feature to adapt the pre-trained input scale to the downstream high-resolution detection input. Step 5: Downstream few-sample target detection fine-tuning: Freeze the backbone network parameters, fine-tune only the classification and regression layers of the detection head, use few-sample labeled samples to complete rapid adaptation, and output the target detection results.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Multi-source SAR data integration. Collect and integrate publicly available SAR datasets, covering X, C, and L bands, with a resolution range of 0.1m to 25m, including single-polarization, dual-polarization, and fully polarization data; Step 1.2: Unified data preprocessing, scaling all SAR images to a resolution of 256×256, calculating the pixel mean and standard deviation of the global dataset, and performing pixel-by-pixel normalization on all images. Step 1.3: Data augmentation strategy configuration. During the pre-training stage, a geometric augmentation method of random horizontal flipping, vertical flipping and 90° multiple rotation is adopted, without introducing color transformation to preserve SAR scattering intensity characteristics. Step 1.4: Dataset partitioning. The preprocessed dataset is divided into a pre-training set and a validation set in an 8:2 ratio for model training and reconstruction performance evaluation.

3. The method according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Encoder structure design. ResNet50 is used as the feature extraction encoder, retaining its complete convolutional layer, batch normalization layer and residual module structure. The number of input channels is adjusted from the standard 3 channels to 1 channel. The encoder output feature dimension is consistent with the output dimension of the backbone network of the downstream ARS-DETR detection framework. Step 2.2: Design an asymmetric lightweight decoder. Construct a convolutional decoder with a parameter size of 1 / 2.4 of the encoder. It is only responsible for mapping the latent representations output by the encoder back to the original image space, thereby reducing the computational cost of pre-training. The decoder input is the feature map output by the encoder, with dimensions (B, 2048, 8, 8), where B is the batch size. The signal processing flow is as follows: Step 2.2.

1. Input projection dimensionality reduction layer: The number of channels is compressed from 2048 to 512 through 1×1 convolution, followed by ReLU activation function, and the output size is (B,512,8,8); Step 2.2.

2. Four-stage progressive upsampling decoding blocks: Four identical upsampling decoding blocks are concatenated. Each decoding block first doubles the feature map size through bilinear interpolation, then performs local feature enhancement through 3×3 convolution, followed by batch normalization layers and ReLU activation functions, while keeping the number of channels constant at 512. After four stages of upsampling, the feature map sizes become 16×16, 32×32, 64×64, and 128×128 respectively. Step 2.2.

3. Output projection and size alignment layer: The number of channels is compressed to 1 through 1×1 convolution to match the single-channel characteristics of SAR, and then the size is restored to 256×256 through bilinear interpolation. Finally, the reconstructed SAR image is output. Step 2.3: Mask reconstruction mechanism design, the input image is divided into 16×16 non-overlapping image blocks, a binary mask matrix is ​​generated to randomly occlude most of the image area, and the encoder only encodes the visible area; Step 2.4: SAR speckle noise modeling. Random multiplicative noise following a Gamma distribution is introduced into the input image to simulate the inherent speckle characteristics of SAR images, forcing the model to learn noise-resistant structural features rather than noise patterns.

4. The method according to claim 3, characterized in that, Step 2.4 includes: Step 2.4.

1. Generate a standard normally distributed random matrix with the same size as the input image. A Gamma distribution with shape parameter k=2 and scale parameter θ=0.5; Step 2.4.

2. Set the noise injection step before the random masking operation. After performing noise injection on the complete input image, generate a binary mask matrix for random occlusion. Step 2.4.

3. According to the formula Perform pixel-wise multiplicative noise injection, where I is the mask for the original input image. The image after noise injection; Step 2.4.

4. Input the image with injected noise into the encoder for feature extraction.

5. The method according to claim 1, characterized in that, Step 3 includes the following sub-steps: Step 3.1: Mask ratio optimization setting, using a mask ratio of 0.6 to balance noise suppression and target structure preservation; Step 3.2: Loss function definition, using mean squared error as the reconstruction loss, only calculating the pixel reconstruction error of the masked region, without constraining the visible region; Step 3.3: Optimizer and learning rate scheduling, using the AdamW optimizer, with an initial learning rate set to 1.5 × 10⁻⁶. -4 A linear warm-up mechanism is introduced for the first 2000 iterations, and a cosine annealing learning rate scheduling strategy is adopted thereafter. Step 3.4: Training stability control, set the gradient clipping threshold to 1.0 to suppress the gradient explosion problem caused by strong scattering points in SAR images, set the batch size to 40, and the total number of training rounds to 200.

6. The method according to claim 1, characterized in that, Step 4 includes the following sub-steps: Step 4.1: Pre-trained weight extraction. After training, discard the decoder part used only for the reconstruction task and extract all weight parameters of the ResNet50 encoder. Step 4.2: Detection model structure adaptation. The extracted encoder weights are fully loaded into the backbone network of the downstream ARS-DETR detection framework. The neck network, Transformer codec, and detection head of the detection framework are all randomly initialized using the Xavier method. Step 4.3: Input scale adaptation. Utilizing the fully convolutional characteristics of ResNet50, the 256×256 input scale during the pre-training stage is directly adapted to the 1024×1024 high-resolution input of the downstream detection task without adjusting the convolution kernel parameters.

7. The method according to claim 1, characterized in that, Step 5 includes the following sub-steps: Step 5.1: Full-supervised training of base classes. The transferred detection model is trained in a fully supervised manner using the 17 base class labeled data of the FAIR-CSAR dataset to optimize the joint parameters of the backbone network and the detection head. Step 5.2: Fine-tuning with a small number of new class samples. Freeze all parameters of the backbone network, Transformer encoder and neck network, and fine-tune only the classification layer and regression layer of the detection head. Use 5-shot, 10-shot or 20-shot of new class labeled samples to complete the rapid adaptation. Step 5.3: Model Inference. Input the SAR image to be detected, and the model outputs the five parameters of the rotated bounding box of the target. The system outputs the final detection results after filtering out predicted bounding boxes with a confidence score below 0.05, along with category labels and confidence scores.

8. The method according to claim 2, characterized in that, The formula for pixel-by-pixel normalization in step 1.2 is: in These are the original image pixel values. These are the normalized pixel values. The global pixel mean. The standard deviation is denoted as .

9. The method according to claim 3, characterized in that, The formula for injecting multiplicative noise with the Gamma distribution in step 2.4 is: in The original input image, Let be a random variable following a Gamma distribution with shape parameter k=2 and scale parameter θ=0.

5. This is the image after noise has been injected.

10. The method according to claim 5, characterized in that, Step 3.2 Mask region mean square error loss: in The total number of pixels being masked. The set of pixel indices for the masked region. The pixel values ​​reconstructed by the model. These are the actual pixel values ​​of the original image. Only the loss in the masked region is calculated; the visible region is not constrained.

Images