A target detection method and system based on SAR imaging and despeckle detection

Abstract

This invention discloses a target detection method and system based on SAR imaging and speckle removal. The method includes: acquiring samples and constructing a training dataset; constructing a SAR imaging module; constructing a speckle removal module; constructing a target detection module, and combining the SAR imaging module and the speckle removal module to obtain a target detection model based on SAR imaging and speckle removal; training the target detection model based on the training dataset and a triple distillation model learning method to obtain a trained target detection model based on SAR imaging and speckle removal. The system includes a data acquisition module, a SAR imaging module, a speckle removal module, a target detection module, and a model training module. Using this invention, integrated processing from SAR image imaging, speckle removal, to target detection can be achieved in real-time and with high precision. This invention can be widely applied in the field of image processing technology.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a target detection method and system based on SAR imaging and despeckle removal. Background Technology

[0002] Synthetic Aperture Radar (SAR) imaging technology mainly includes frequency domain processing, time domain processing, and deep learning algorithms. In frequency domain processing, the Range-Doppler (RD) algorithm achieves high-resolution imaging by separating range and azimuth dimensions, but it does not consider range cell movement, limiting image quality. Therefore, improved algorithms have been proposed to partially address this issue, such as the Secondary Range Compression (SRC) algorithm and the Frequency Modulation Scaling (CS) algorithm. However, these improved algorithms have limited effectiveness in large oblique-view scenarios. Further, the ωK algorithm is proposed, which effectively corrects range migration through Stolt interpolation and is applicable to various oblique viewpoints. In time domain processing, the Back Projection (BPA) algorithm achieves accurate imaging by compensating for the propagation delay of pulse signals point-by-point, but it suffers from high computational cost and low efficiency. In the application of deep learning technology, autofocusing imaging based on recurrent neural networks (RNNs) improves imaging accuracy and robustness through automatic feature learning. Furthermore, fast back projection based on convolutional neural networks (CNNs) and the SAR-GMTIm framework for moving targets significantly improve imaging speed and quality.

[0003] In the study of speckle removal from synthetic aperture radar (SAR) images, the main methods employed include spatial domain filtering, transform domain filtering, and deep learning. In spatial domain filtering, filters such as Lee, Kuan, and Frost are used to reduce speckle by calculating the weighted average of pixel intensity within a local window. While this method performs well in uniform regions, it produces poor edge-level results, and its effectiveness decreases as the window size increases. In transform domain filtering, wavelet domain filtering improves denoising through logarithmic transformation and thresholding after wavelet decomposition, but this may introduce distortion. The Bayesian wavelet thresholding algorithm improves speckle removal by assuming a Gaussian distribution. In deep learning techniques, PSD, ID-CNN, and Speckle2Void achieve denoising by learning the mapping relationship between noise and sharp images. SAR-CNN utilizes homomorphic transformation and residual networks to improve noise processing. GAN combined with RED-NET significantly improves denoising efficiency and quality.

[0004] In ship detection using synthetic aperture radar (SAR) images, traditional methods mainly include contrast-based methods, template matching, statistical analysis, and deep learning. Constant False Alarm Rate (CFAR) is the most commonly used SAR image target detection technique, achieving good performance in some simple scenarios. However, its performance largely depends on the construction of the sea clutter statistical model and the parameter estimation of the model. Therefore, considering the inhomogeneity of clutter and various clutter models, a method based on heterogeneous distributions to fit variable ocean conditions, such as alpha-stable and gamma distributions, is proposed.

[0005] However, with the continuous development of technology, these algorithms can no longer meet modern needs and are difficult to apply to various complex synthetic aperture radar (SAR) image scenarios. Previous research has separated SAR image despecization and target detection, a process that requires multiple data transmissions and independent processing, resulting in low efficiency. Furthermore, the lack of information exchange and utilization between different stages limits the processing results to localized information at each stage, leading to error accumulation in downstream tasks and failing to provide fast and high-precision technology. Therefore, how to achieve integrated processing from SAR image imaging and despecization to target detection in real-time and with high precision has become a challenging problem in the field of SAR technology research. Summary of the Invention

[0006] To address the aforementioned technical problems, the present invention aims to provide a target detection method and system based on SAR imaging and despectics, which can achieve integrated processing from SAR image imaging, despectics removal to target detection tasks in real time and with high precision.

[0007] The first technical solution adopted in this invention is: a target detection method based on SAR imaging and despeckle detection, comprising the following steps:

[0008] Samples were collected from high-resolution remote sensing images to construct a training dataset;

[0009] A SAR imaging module is constructed based on the multi-scale window Swin Transformer module;

[0010] A speckle removal module is constructed based on the multi-scale window Swin Transformer module;

[0011] A target detection module is constructed and combined with the SAR imaging module and the despecciation module to obtain a target detection model based on SAR imaging and despecciation.

[0012] The target detection model based on SAR imaging and despeciation is trained using the training dataset and triple distillation model learning method to obtain the trained target detection model based on SAR imaging and despeciation.

[0013] Furthermore, the training dataset includes SAR images, despeculiarized SAR images, and target detection labels.

[0014] Furthermore, the SAR imaging module includes convolutional layers, residual blocks, and a multi-scale window Swing Transformer module, wherein:

[0015] The multi-scale window Swin Transformer module is used to learn global and local features of SAR images;

[0016] The residual block introduces skip connections to address the potential gradient vanishing problem in deep neural networks and further extract feature information from the image.

[0017] The convolutional layer is used to remove speckle noise from SAR images.

[0018] Furthermore, the multi-scale window Swing Transformer module comprises a first sub-unit and a second sub-unit, wherein:

[0019] The first sub-unit includes a Layer Norm layer, a local feature extraction module, and a first multi-window attention module;

[0020] The second subunit includes a Layer Norm layer, a local feature extraction module, and a second multi-window attention module;

[0021] The first multi-window attention module adopts a multi-head self-attention mechanism for windows;

[0022] The second multi-window attention module adopts a sliding window multi-head self-attention mechanism.

[0023] Furthermore, the first multi-window attention module and the second multi-window attention module use multiple windows of different sizes to perform self-attention calculations.

[0024] Furthermore, the step of training the target detection model based on SAR imaging and despeculiaration using the training dataset and the triple distillation model learning method to obtain the trained target detection model based on SAR imaging and despeculiaration specifically includes:

[0025] The existing trained SAR imaging model, despeculiarization model, and target detection model are respectively used as the first teacher model, the second teacher model, and the third teacher model.

[0026] The training dataset is input into the SAR imaging module and the first teacher model respectively to obtain the first prediction result and the first teacher model prediction result;

[0027] A first distillation loss function is constructed based on the first prediction result and the first teacher model prediction result;

[0028] An imaging loss function is constructed based on the first prediction result and the real image;

[0029] The training dataset is input into the speckle removal module and the second teacher model respectively to obtain the second prediction result and the prediction result of the second teacher model;

[0030] A second distillation loss function is constructed based on the second prediction result and the second teacher model prediction result;

[0031] Based on the second prediction result and the real image, a speckle removal loss function is constructed;

[0032] The training dataset is input into the object detection module and the third teacher model respectively to obtain the third prediction result and the third teacher model prediction result;

[0033] A third distillation loss function is constructed based on the third prediction result and the third teacher model prediction result;

[0034] A detection loss function is constructed based on the third prediction result and the target detection label;

[0035] An additional loss term is introduced, and weights are assigned to the first distillation loss function, the imaging loss function, the second distillation loss function, the despeckle loss function, the third distillation loss function, and the detection loss function to obtain a global balanced loss function;

[0036] The parameters of the target detection model based on SAR imaging and despeculiarization are adjusted based on the global balance loss function to obtain the trained target detection model based on SAR imaging and despeculiarization.

[0037] The second technical solution adopted in this invention is: a target detection system based on SAR imaging and despeckle detection, comprising:

[0038] The data acquisition module is used to collect samples from high-resolution remote sensing images and build a training dataset.

[0039] The SAR imaging module is used to perform imaging processing on the input SAR echo signal;

[0040] Despeckle module; used to perform despeckle processing on the imaging results output by the SAR imaging module;

[0041] The target detection module is used to perform image detection on the despeckled SAR image;

[0042] The model training module trains the target detection model based on SAR imaging and despeciation using the training dataset and the triple distillation model learning method, resulting in the trained target detection model based on SAR imaging and despeciation.

[0043] This invention also proposes a target detection device based on SAR imaging and despeckle detection, comprising:

[0044] At least one processor;

[0045] At least one memory for storing at least one program;

[0046] When the at least one program is executed by the at least one processor, the at least one processor implements a target detection method based on SAR imaging and despeckle as described above.

[0047] The beneficial effects of the method and system of this invention are as follows: This invention proposes a multi-scale window SwinTransformer module, and uses this module to construct a SAR imaging module and a despecciation module, which can capture global and local relationships in the image, thereby obtaining richer information and helping to improve the model's ability to represent images and enhance image reconstruction capabilities. By combining the target detection module, SAR imaging module, and despecciation module, a target detection model based on SAR imaging and despecciation is constructed, and the triple distillation model training method is used to train the model, improving training accuracy and obtaining the trained target detection model based on SAR imaging and despecciation. The trained target detection model based on SAR imaging and despecciation can realize the integrated processing of SAR image imaging, despecciation, and target detection tasks in real time and with high accuracy. Attached Figure Description

[0048] Figure 1 This is a flowchart of the steps of a target detection method based on SAR imaging and despeckle removal according to the present invention;

[0049] Figure 2 This is a structural block diagram of a target detection system based on SAR imaging and despeckle detection according to the present invention;

[0050] Figure 3 This is a structural diagram of a target detection model based on SAR imaging and despeculiaration, which is a target detection method based on SAR imaging and despeculiaration according to the present invention.

[0051] Figure 4This is a structural diagram of the imaging module of a target detection method based on SAR imaging and despeckle detection according to the present invention.

[0052] Figure 5 This is a structural diagram of the multi-scale window SwinTransformer module of the present invention, which is a target detection method based on SAR imaging and despeckle detection.

[0053] Figure 6 This is a structural diagram of the local feature extraction module of a target detection method based on SAR imaging and despeckle detection according to the present invention;

[0054] Figure 7 This is a schematic diagram of the triple distillation model learning and loss calculation of a target detection method based on SAR imaging and despeckle detection according to the present invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0056] Reference Figure 1 This invention provides a target detection method based on SAR imaging and despeckle removal, the method comprising the following steps:

[0057] S1. Collect samples from high-resolution remote sensing images and construct a training dataset;

[0058] Specifically, the training dataset includes SAR images, despecced SAR images, and target detection labels.

[0059] S2. Construct a SAR imaging module based on the multi-scale window Swin Transformer module;

[0060] Specifically, refer to Figure 4 The SAR imaging module includes two multi-scale window Swin Transformer (MSW Swin Transformer) modules, a convolutional layer (Conv), and a residual block.

[0061] The implementation process of the SAR imaging module specifically includes: First, learning global and local features of the SAR image through two multi-scale window Swin Transformer modules. The multi-scale window Swin Transformer modules capture global and local relationships in the image through a multi-scale window self-attention mechanism, thereby acquiring richer information. This helps improve the model's ability to represent the image and enhances image reconstruction capabilities. Then, four residual blocks are used to further extract feature information from the image. The residual blocks, by introducing skip connections, solve the potential gradient vanishing problem in deep neural networks, thus ensuring the stability and efficiency of the network during deep training; simultaneously, the residual blocks can accelerate convergence during training and help the network learn more complex and abstract feature representations. Finally, two convolutional layers are used to remove speckle noise from the SAR image.

[0062] The self-attention mechanism in Transformers helps the model capture global and local relationships in images, enabling Transformers to perform well in tasks requiring an understanding of the global structure of an image. Since Transformer models can face high computational complexity when processing large-scale images, the Swing Transformer was proposed. The Swing Transformer, by introducing a multi-window self-attention mechanism, reduces computational complexity to some extent and improves the model's computational efficiency. The Swing Transformer module can utilize self-attention to capture global and local contextual information in an image, thus helping to distinguish noise from signals and better understand the structure and content of the image. However, the fixed-window self-attention mechanism of the Swing Transformer module suffers from local and boundary effects, easily segmenting a single target into multiple windows, leading to inaccurate self-attention calculations. Therefore, this invention proposes a multi-scale window Swing Transformer module based on the Swing Transformer module. This module introduces a multi-window attention module (MWA). Smaller windows help the model focus more on fine structures and details in the image, while larger windows are better suited for acquiring global contextual information.

[0063] Reference Figure 5 The multi-scale window Swing Transformer module comprises a first sub-unit and a second sub-unit, wherein:

[0064] The output of the first sub-unit is the input of the second sub-unit;

[0065] The first sub-unit includes a Layer Norm layer, a local feature extraction module, and a first multi-window attention module;

[0066] The second subunit includes a Layer Norm layer, a local feature extraction module, and a second multi-window attention module;

[0067] The first multi-window attention module adopts a multi-head self-attention mechanism for windows;

[0068] The second multi-window attention module adopts a sliding window multi-head self-attention mechanism.

[0069] Both the first and second multi-window attention modules contain three branches, each employing three different window sizes: 4, 8, and 16. In the subsequent SAR image despeccing task, acquiring more contextual information is crucial for improving model performance. Therefore, this invention designs three different window sizes and performs self-attention calculation within each window. Then, the similarity probability maps output by the self-attention mechanisms of the three different window sizes are summed and normalized.

[0070] Reference Figure 6 In the multi-scale window Swin Transformer module of this invention, a Localized Feature Extraction (LFE) module is used instead of the MLP layer. The LFE module consists of two 1*1 convolutional layers and a 3*3 convolutional layer with a dilation rate r=2. The original MLP layer relies on linear transformations and has weak abstraction capabilities for features. By introducing the LFE module to replace the MLP layer, the network can better extract local features of the image in the task of removing speckle noise, exhibiting stronger image detail and texture capabilities.

[0071] S3. Construct a speckle removal module based on the multi-scale window Swin Transformer module;

[0072] Specifically, the despeculiarization module is used to despeculate the imaging results output by the SAR imaging module, and its structure is the same as that of the SAR imaging module in step S2.

[0073] S4. Construct a target detection module and combine it with the SAR imaging module and the despeccation module to obtain a target detection model based on SAR imaging and despeccation.

[0074] Specifically, refer to Figure 3By combining the coding module, SAR imaging module, speckle module, and target detection module into a single model, a target detection model based on SAR imaging and speckle removal is obtained.

[0075] S5. The target detection model based on SAR imaging and despeciation is trained based on the training dataset and the triple distillation model learning method to obtain the trained target detection model based on SAR imaging and despeciation.

[0076] Specifically, refer to Figure 7 The triple distillation model learning method proposed in this invention is a training method based on the "teacher-student network concept," which is widely used in industry due to its simplicity and effectiveness. By distilling the feature representation knowledge learned by a complex, high-learning-capability network, and then passing it to a network with fewer parameters and weaker learning ability, distillation can provide soft label information that the student model cannot learn from the ground truth labels. This includes inter-class information and feature representation knowledge that the student network cannot learn but the teacher network can, thus generally improving the accuracy of the student network. In a specific embodiment of this invention, existing trained SAR imaging models, despeckle models, and target detection models are used as the first, second, and third teacher models, respectively, to guide the learning and training of the SAR imaging module, despeckle module, and target detection module. The specific process is as follows:

[0077] The training dataset is input into the SAR imaging module and the first teacher model respectively to obtain the first prediction result and the first teacher model prediction result;

[0078] A first distillation loss function is constructed based on the first prediction result and the first teacher model prediction result;

[0079] An imaging loss function is constructed based on the first prediction result and the real image;

[0080] The training dataset is input into the speckle removal module and the second teacher model respectively to obtain the second prediction result and the prediction result of the second teacher model;

[0081] A second distillation loss function is constructed based on the second prediction result and the second teacher model prediction result;

[0082] Based on the second prediction result and the real image, a speckle removal loss function is constructed;

[0083] The training dataset is input into the object detection module and the third teacher model respectively to obtain the third prediction result and the third teacher model prediction result;

[0084] A third distillation loss function is constructed based on the third prediction result and the third teacher model prediction result;

[0085] A detection loss function is constructed based on the third prediction result and the target detection label.

[0086] The first distillation loss function, the imaging loss function, the second distillation loss function, the despeculiarization loss function, the third distillation loss function, and the detection loss function together constitute the overall loss function of the target detection model based on SAR imaging and despeculiarization. Specifically, the first distillation loss function guides the SAR imaging module to learn from the first teacher model; the second distillation loss function guides the despeculiarization module to learn from the second teacher model; and the third distillation loss function guides the target detection module to learn from the third teacher model. The distillation loss function measures the difference between the student model's prediction results and the teacher model's prediction results, aiming to enable the student model to imitate the teacher model's prediction behavior as closely as possible. The imaging loss function, despeculiarization loss function, and detection loss function are corresponding loss functions specifically designed for the three tasks of SAR imaging, despeculiarization, and target detection, mainly used to measure the model's prediction accuracy on real labels. In this specific embodiment, a pixel-level mean squared error loss function is used as the imaging loss function and the despeculiarization loss function to measure the difference between the generated image and the real image; a cross-entropy loss function or a target detection-specific loss function is used as the detection loss function to measure the model's prediction accuracy for the target.

[0087] To balance the loss functions of various tasks within the overall deep learning model, this invention assigns weights to the first distillation loss function, the imaging loss function, the second distillation loss function, the despeckle loss function, the third distillation loss function, and the detection loss function, and then performs a weighted summation. By adjusting the weights of different tasks, the training importance among them is balanced. The weight coefficients can be adjusted according to actual conditions to achieve balance and optimization for each task. Typically, the weight coefficients can be determined using techniques such as cross-validation to ensure the model achieves optimal performance on each task. Furthermore, to further enhance the role of the global balancing loss, specific embodiments of this invention incorporate the correlation and mutual influence between tasks, introducing a correlation penalty term or a cross-loss term between tasks as additional loss terms. These additional loss terms help the model better understand the relationships between tasks, thereby achieving better balance and overall performance.

[0088] Reference Figure 2 This invention provides a target detection system based on SAR imaging and despeckle detection, comprising:

[0089] The data acquisition module is used to collect samples from high-resolution remote sensing images and build a training dataset.

[0090] The SAR imaging module is used to perform imaging processing on the input SAR echo signal;

[0091] Despeckle module; used to perform despeckle processing on the imaging results output by the SAR imaging module;

[0092] The target detection module is used to perform image detection on the despeckled SAR image;

[0093] The model training module trains the target detection model based on SAR imaging and despeciation using the training dataset and the triple distillation model learning method, resulting in the trained target detection model based on SAR imaging and despeciation.

[0094] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system 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.

[0095] A target detection device based on SAR imaging and despeckle detection:

[0096] At least one processor;

[0097] At least one memory for storing at least one program;

[0098] When the at least one program is executed by the at least one processor, the at least one processor implements a target detection method based on SAR imaging and despeckle as described above.

[0099] 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.

[0100] A storage medium storing processor-executable instructions, which, when executed by a processor, are used to implement a target detection method based on SAR imaging and despeckle removal as described above.

[0101] 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.

[0102] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A target detection method based on SAR imaging and despeckle detection, characterized in that, The following steps are involved: Samples were collected from high-resolution remote sensing images to construct a training dataset; A SAR imaging module is constructed based on the multi-scale window Swin Transformer module; A speckle removal module is constructed based on the multi-scale window Swin Transformer module; A target detection module is constructed and combined with the SAR imaging module and the despecciation module to obtain a target detection model based on SAR imaging and despecciation. The target detection model based on SAR imaging and despeciation is trained using the training dataset and triple distillation model learning method to obtain the trained target detection model based on SAR imaging and despeciation. The multi-scale window Swin Transformer module comprises a first sub-unit and a second sub-unit, wherein: The first sub-unit includes a Layer Norm layer, a local feature extraction module, and a first multi-window attention module; The second subunit includes a Layer Norm layer, a local feature extraction module, and a second multi-window attention module; The first multi-window attention module adopts a multi-head self-attention mechanism for windows; The second multi-window attention module adopts a sliding window multi-head self-attention mechanism; The first multi-window attention module and the second multi-window attention module use multiple windows of different sizes to perform self-attention calculation; The local feature extraction module consists of two 1x1 convolutional layers and a dilatation layer. r Composed of 3x3 convolutional layers with a value of 2; in, The existing trained SAR imaging model, despeculiarization model, and target detection model are respectively used as the first teacher model, the second teacher model, and the third teacher model. The training dataset is input into the SAR imaging module and the first teacher model respectively to obtain the first prediction result and the first teacher model prediction result; A first distillation loss function is constructed based on the first prediction result and the first teacher model prediction result; An imaging loss function is constructed based on the first prediction result and the real image; The training dataset is input into the speckle removal module and the second teacher model respectively to obtain the second prediction result and the prediction result of the second teacher model; A second distillation loss function is constructed based on the second prediction result and the second teacher model prediction result; Based on the second prediction result and the real image, a speckle removal loss function is constructed; The training dataset is input into the object detection module and the third teacher model respectively to obtain the third prediction result and the third teacher model prediction result; A third distillation loss function is constructed based on the third prediction result and the third teacher model prediction result; A detection loss function is constructed based on the third prediction result and the target detection label; An additional loss term is introduced, and weights are assigned to the first distillation loss function, the imaging loss function, the second distillation loss function, the despeckle loss function, the third distillation loss function, and the detection loss function to obtain a global balanced loss function; The parameters of the target detection model based on SAR imaging and despeculiarization are adjusted based on the global balance loss function to obtain the trained target detection model based on SAR imaging and despeculiarization.

2. The target detection method based on SAR imaging and despeckle removal according to claim 1, characterized in that, The training dataset includes SAR images, despecimenized SAR images, and target detection labels.

3. The target detection method based on SAR imaging and despeckle removal according to claim 1, characterized in that, The SAR imaging module includes convolutional layers, residual blocks, and a multi-scale window Swing Transformer module, wherein: The multi-scale window Swin Transformer module is used to learn global and local features of SAR images; The residual block introduces skip connections to address the potential gradient vanishing problem in deep neural networks and further extract feature information from the image. The convolutional layer is used to remove speckle noise from SAR images.

4. A target detection system based on SAR imaging and despeckle detection, characterized in that, A method for performing a target detection method based on SAR imaging and despeckle as described in claim 1 includes: The data acquisition module is used to collect samples from high-resolution remote sensing images and build a training dataset. The SAR imaging module is used to perform imaging processing on the input SAR echo signal; The despeculiarization module is used to perform despeculiarization processing on the imaging results output by the SAR imaging module. The target detection module is used to perform image detection on the despeckled SAR image; The model training module trains the target detection model based on SAR imaging and despeciation using the training dataset and the triple distillation model learning method, resulting in the trained target detection model based on SAR imaging and despeciation.

5. A target detection device based on SAR imaging and despeckle detection, 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 a target detection method based on SAR imaging and despeckle as described in any one of claims 1-3.

Images