Self-supervised deep learning sidelobe and clutter suppression method for SAR image ship target
Through self-supervised deep learning combined with tone mapping and clustering algorithm, the problem of sidelobe and noise suppression technology sacrificing detailed information in SAR images is solved, and efficient denoising and sidelobe suppression are achieved, which significantly improves image quality and subsequent processing. accuracy.
Patent Information
- Application Number
- CN202411030278.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-07-30
AI Technical Summary
Existing sidelobe and noise suppression technologies will sacrifice the detailed information of the image while reducing noise, affecting image resolution, especially in SAR image processing.
The self-supervised deep learning method is used to combine tone mapping technology and clustering algorithm to preprocess and target rectangle profile extraction on SAR images, suppress side lobes and clutter through weighting functions and bilateral filtering algorithms, and update the deep learning model with loss functions. Denoising and sidelobe suppression are achieved.
Effectively suppress side lobes and clutter, preserve ship target details, improve image quality and resolution, reduce computing resource requirements, and is suitable for practical application scenarios.
Smart Images

Figure CN118918030A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to a self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets. Background Art
[0002] Synthetic aperture radar (SAR) remote sensing technology has become an important tool for ocean monitoring because it can penetrate clouds and rain curtains and effectively monitor ships at sea under various weather conditions. However, the imaging characteristics of SAR images, such as the sidelobe effect of strong scattering points and the accompanying speckle noise, seriously affect the image quality. These problems not only reduce the usability of the image, but also pose challenges to subsequent processing tasks such as image segmentation, target detection and recognition. Therefore, exploring effective speckle noise suppression methods is of great significance to improving the quality of SAR images.
[0003] At present, traditional sidelobe and noise suppression technologies often sacrifice image detail information while reducing noise, thereby affecting image resolution. Summary of the invention
[0004] The purpose of the present invention is to propose a self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets in order to address the problem that existing sidelobe and noise suppression technologies often sacrifice image detail information while reducing noise, thereby affecting image resolution.
[0005] The technical solution adopted by the present invention to solve the above technical problems is:
[0006] A self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets, including:
[0007] Step 1: Acquire a SAR image containing a ship target;
[0008] Step 2: Process the SAR image using tone mapping technology;
[0009] Step 3: Use clustering algorithm to extract the target rectangle outline from the image processed by tone mapping technology;
[0010] Step 4: Use a weighted function to filter and cluster each extracted target rectangular outline to obtain the outline of the ship target, and map the outline of the ship target to the SAR image in step 1 to obtain a mapped image;
[0011] Step 5: Use the SAR image in step 1 and the image after mapping to obtain the loss function, and update the supervised deep learning model through the loss function until the loss function converges to obtain the image after denoising and sidelobe suppression.
[0012] Furthermore, the SAR image containing the ship target in step 1 is a preprocessed image, and the preprocessing includes complex data conversion, multi-view processing, filtering processing and geocoding.
[0013] Furthermore, the specific steps of step 2 are:
[0014] Step 21: Calculate the logarithmic mean Y of the SAR image containing the ship target av :
[0015] Step 22: Use the brightness value and logarithmic mean of the pixels in the image to get the key value;
[0016] Step 2 and 3: Use the key value to partition the SAR image. The specific steps are as follows:
[0017] Use the key value to get the threshold L of the low-light area of the brightness image T and the threshold L of the high illumination area of the brightness image H , and the pixel brightness in the SAR image is less than L T The pixel brightness is greater than L H The pixel brightness is between L T and L H Those in between are classified as medium illumination areas;
[0018] Step 24: The partitioned SAR image is compressed using an adaptive logarithmic method to obtain the display brightness after tone mapping.
[0019] Furthermore, the logarithmic mean Y of the SAR image containing the ship target av It is expressed as:
[0020]
[0021] Among them, Y(x) represents the brightness channel of the image converted into XYZ space, W is the total number of pixels in the image, δ represents a positive number, and δ is 10 -4 , x represents the pixel position index in the image.
[0022] Furthermore, the key value is expressed as:
[0023]
[0024] Among them, Y max Indicates the maximum brightness value of the pixel, Y min Indicates the minimum brightness value of a pixel.
[0025] Furthermore, the L T and L H It is expressed as:
[0026] L T =L max -(0.9+0.1key)(L max -L min )
[0027] L H =L min +[0.6+0.4×(1-key)](L max -L min )
[0028] Among them, L max and L min The images are normalized to the maximum and minimum values between 0 and 1, respectively.
[0029] Furthermore, the display brightness after tone mapping is expressed as:
[0030]
[0031] Among them, L i (x, y) represents the display brightness after tone mapping, L w (x,y) and L w,max They represent the original brightness and the maximum value of the original brightness respectively. q and k are constants. i ∈[1,∞],k i ∈[1,∞]; i=T,W,H; j=1,2, when j=1, S1=10, j=2, S2=2, S represents the base of the logarithmic mapping.
[0032] Furthermore, the clustering algorithm in step 3 is a K-means clustering algorithm, which is specifically expressed as follows:
[0033]
[0034] Among them, x i represents the i-th sample, that is, the i-th pixel in the image, and j represents x i The cluster, p j represents the center point corresponding to the cluster, and n represents the total number of samples.
[0035] Furthermore, the specific steps of step 4 are:
[0036] Step 41: Use bilateral filtering algorithm to process the extracted target rectangular contour;
[0037] Step 42: Cluster the bilaterally filtered image to obtain a refined target contour;
[0038] Step 43: Map the refined target contour back to the original SAR image to obtain a mapped image.
[0039] Furthermore, the bilateral filtering algorithm is expressed as:
[0040]
[0041] Among them, w(i,j,k,l) represents the weight function in the bilateral filtering algorithm, d(i,j,k,l) represents the weighted combination of the values of the neighborhood pixels, r(i,j,k,l) represents the range template coefficient, (k,l) represents the center coordinates of the template window, (i,j) represents the coordinates of the other coefficients of the template window, σ d represents the standard deviation of the Gaussian function, f(i,j) represents the pixel value of the image at point (i,j), f(k,l) represents the pixel value of the image at point (k,l), σ r Represents the standard deviation of the Gaussian function.
[0042] The beneficial effects of the present invention are:
[0043] The present application can suppress side lobes and clutter while retaining the details of the ship target, thus avoiding the impact on image resolution.
[0044] This application achieves efficient denoising and sidelobe suppression of SAR images through a self-supervised deep learning framework combined with tone mapping technology and clustering algorithms, without relying on a large amount of annotated data or pre-known noise distribution. This method can automatically learn and retain the details of ship targets, while compressing the high dynamic range and improving image contrast and detail visibility, significantly improving the quality of SAR images and the accuracy of subsequent processing tasks. In addition, this application reduces the demand for computing resources and improves processing efficiency, making it more suitable for practical application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 A flow chart of the steps implemented for this application;
[0046] Figure 2 This is the result image obtained by tone mapping, clustering and weighting function;
[0047] Figure 3 Comparison chart of experimental results of the proposed self-supervised deep learning method. DETAILED DESCRIPTION
[0048] It should be particularly noted that, in the absence of conflict, the various embodiments disclosed in this application can be combined with each other.
[0049] Specific implementation method 1: refer to Figure 1 Specifically describing this embodiment, the self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets described in this embodiment includes:
[0050] Step 1: Acquire a SAR image containing a ship target;
[0051] Step 2: Process the SAR image using tone mapping technology;
[0052] Step 3: Use clustering algorithm to extract the target rectangle outline from the image processed by tone mapping technology;
[0053] Step 4: Use a weighted function to filter and cluster each extracted target rectangular outline to obtain the outline of the ship target, and map the outline of the ship target to the SAR image in step 1 to obtain a mapped image;
[0054] Step 5: Using the SAR image in step 1 and the image after mapping, a loss function is obtained through a deep learning method, and the supervised deep learning model is updated through the loss function until the loss function converges to obtain an image after denoising and sidelobe suppression.
[0055] This application aims to enable the model to achieve high-quality denoising and sidelobe suppression of SAR images at a low cost without any training data or prior knowledge of noise distribution through a self-supervised learning framework, and automatically learn the ability to suppress sidelobes and clutter while retaining the details of ship targets. This method uses the inherent characteristics of SAR images and achieves effective suppression of sidelobes and clutter through a set loss function and network structure.
[0056] In order to achieve the above purpose, the technical solution adopted in this application includes the following steps:
[0057] 1. Acquire a SAR image containing a ship target;
[0058] Step S1, extract the map data and image data containing the ship target from the existing open source data, and perform preprocessing on the map data and image data including image registration, filtering, etc. Then, extract a SAR image containing the ship target with a size of 500×500 pixels from the preprocessed image as the input of the network.
[0059] 2. Preprocess SAR images using tone mapping techniques to adapt to their high dynamic range characteristics, including:
[0060] Logarithmic tone mapping technique is used to preprocess SAR images to effectively adapt to their high dynamic range characteristics and improve image contrast and detail visibility.
[0061] Step S2, using the adaptive logarithmic tone mapping method to process the SAR image slices, adjust the brightness value of the image to effectively adapt to its high dynamic range characteristics, and improve the image contrast and detail visibility. The details are as follows:
[0062] Step S21, calculate the logarithmic mean of the SAR image:
[0063]
[0064] Where: Y(x) represents the brightness channel of the image converted to XYZ space; W is the total number of pixels in the image; δ is a small positive number, in this example δ is 10 -4 ;
[0065] Step S22, use the logarithmic average of the image brightness value Y as the key value:
[0066]
[0067] Among them, Y max is the maximum brightness value; Y min is the minimum brightness value.
[0068] Step S23, partitioning the SAR image according to the key value as a threshold:
[0069] L T =L max -(0.9+0.1key)(L max -L min )
[0070] L H =L min +[0.6+0.4×(1-key)](L max -L min )
[0071] Where: L T is the threshold of the low illumination area of the brightness image; L H is the threshold of the high illumination area of the brightness image; L max and L min are the maximum and minimum values of the image normalized to between 0 and 1. The image brightness is less than L T When it is low illumination area; when it is greater than L H When L T and L H The brightness area in between is the medium illumination area.
[0072] Step S24, using an adaptive logarithmic method to achieve high dynamic range compression processing:
[0073]
[0074] Where: L i (x, y) is the display brightness after tone mapping; L w (x,y) and Lw,max They represent the original brightness and the maximum value of the original brightness respectively; q i ∈[1,∞],k i ∈[1,∞]; when i=T,W,H=1,2,j=1, S1=10, j=2, S2=2, S is the base of the logarithmic mapping; q and k are constants.
[0075] 3. Use clustering algorithm to extract the target rectangular contour from the preprocessed image;
[0076] The DBSCAN clustering algorithm (afterwards binarization) is applied to process the image data, and then the minimum area rectangle that can enclose the contour is calculated using the center point coordinates of the rectangle, the width and height of the rectangle, and the rotation angle of the rectangle to obtain the target area and background area in the SAR image.
[0077] Step S3, using K-means clustering algorithm to extract the target rectangular contour of the preprocessed image, as follows: determine an initial cluster center for each cluster, calculate the objective function of the cluster center, assign samples to the center vector closest to them, and reduce the objective function value.
[0078]
[0079] In the formula, x i represents the i-th sample, j represents x i Cluster, p j Indicates the center point corresponding to the cluster, and n indicates the total number of samples. According to the minimum distance principle, the samples are assigned to the nearest cluster, and the sample mean in each cluster is used as the new cluster center. The cluster center is repeatedly updated until convergence, and finally K clusters are obtained. In this example, K is set to 2.
[0080] 4. Use weighted functions to accurately filter and cluster the extracted contours to distinguish the background area from the target area, obtain the contours of the ship target, and further map the extracted contours of the ship target to the original SAR image to obtain the mapped image;
[0081] Step S4, using a weighted function to perform precise filtering and clustering processing on the extracted contours to distinguish the background area and the target area, and obtain a contour similar to the ship target, and further map the extracted contour similar to the ship target to the original SAR image, in order to obtain Figure 2 To achieve the results shown, the following specific steps are required:
[0082] Step S41: For each clustered contour, a bilateral filtering algorithm is applied to perform edge-preserving denoising processing to maintain the clarity of the target contour while removing speckle noise and subtle clutter in the image. The bilateral filtering algorithm is defined as:
[0083]
[0084] Where: d(i,j,k,l) is the weighted combination of the values of the neighborhood pixels; r(i,j,k,l) is the range template coefficient; (k,l) is the center coordinate of the template window; (i,j) is the coordinate of the other coefficients of the template window; σ d is the standard deviation of the Gaussian function; the function f(x,y) represents the pixel value of the image at point (x,y); σ r is the standard deviation of the Gaussian function.
[0085] Step S42: clustering is performed again on the bilaterally filtered image, and this clustering is intended to further refine the separation of the target from the background, especially in the target edge area. Through this step, the ship target and the side lobes and clutter around it can be distinguished more accurately.
[0086] Step S43: Mapping the refined target contour back to the original SAR image. This step ensures the precise positioning of the target contour by comparing the contour positions in the original image and the pre-processed image, providing accurate target information for sidelobe and clutter suppression.
[0087] 5. Using a self-supervised deep learning model, the network is updated by comparing the error between the processed result and the high dynamic range input image to optimize the denoising and sidelobe suppression effects, including;
[0088] The original image is fed into the convolutional neural network, and after the aforementioned preprocessing steps, the network outputs a mapped image. This image is compared with the original SAR image, and the resulting error is used to guide the optimizer to update the parameters of each layer of the network through the loss function. This process is repeated until the error between the network output and the original noise-free image is minimized, at which point the network training is completed, and the final output is an image that has been denoised and sidelobe suppressed.
[0089] Step S5: Use the UNet network as a self-supervised deep learning model to suppress side lobes and clutter. The UNet model, with its powerful feature extraction capability and effective context information transmission mechanism, can effectively suppress side lobes and clutter without losing target details.
[0090] Step S51: The preprocessed SAR image and the image mapping the target contour are used as inputs of the UNet model, and the model automatically adjusts its parameters through self-supervised learning to minimize the error between the output image and the high dynamic range input image.
[0091] Step S52: Use mean square error (MSE) as the loss function to quantify the difference between the model output and the target image. Optimize the parameters of each layer of the model through the back propagation algorithm to improve the suppression effect of side lobes and clutter.
[0092] Step S53: Repeat steps S51 and S52 until the loss of the model on the validation set no longer decreases significantly, indicating that the model has converged. At this point, the model can effectively denoise and suppress sidelobe in SAR images while retaining important details of the ship target. The results are as follows: Figure 3 shown.
[0093] The overall structure of the model for self-supervised training is UNet. The feature extraction part of UNet uses multiple convolutional layers and pooling layers to gradually extract high-level features in the image, while retaining low-level feature information through jump connections. In self-supervised training, the original image and the preprocessed image are used as the input of the UNet network model, and the denoised and sidelobe suppressed image is used as the output of the model. The loss function uses the mean square error (MSE) loss to quantify the difference between the output image and the target image. The loss function value is obtained through forward propagation, and then the back propagation algorithm is used to optimize the parameters in each convolution layer of the model. Repeat the above steps until the loss function value no longer decreases, indicating that the model has converged. In this way, the UNet model can effectively perform denoising and sidelobe suppression of SAR images through self-supervised learning without external labeled data.
[0094] It should be noted that the specific implementation is only an explanation and description of the technical solution of the present invention, and cannot be used to limit the scope of protection of the rights. Any partial changes made according to the claims and description of the present invention should still fall within the scope of protection of the present invention.
Claims
1. A self-supervised deep learning sidelobe and clutter suppression method for ship targets in SAR images, characterized by include: Step 1: Acquire a SAR image containing a ship target; Step 2: Process the SAR image using tone mapping technology; Step 3: Use clustering algorithm to extract the target rectangle outline from the image processed by tone mapping technology; Step 4: Use a weighted function to filter and cluster each extracted target rectangular outline to obtain the outline of the ship target, and map the outline of the ship target to the SAR image in step 1 to obtain a mapped image; Step 5: Use the SAR image in step 1 and the image after mapping to obtain the loss function, and update the supervised deep learning model through the loss function until the loss function converges to obtain the image after denoising and sidelobe suppression.
2. The self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets according to claim 1, characterized in that The SAR image containing the ship target in step 1 is a pre-processed image, and the pre-processing includes complex data conversion, multi-view processing, filtering processing and geocoding.
3. The self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets according to claim 1, characterized in that The specific steps of step 2 are: Step 21: Calculate the logarithmic mean Y of the SAR image containing the ship target av : Step 22: Use the brightness value and logarithmic mean of the pixels in the image to get the key value; Step 2 and 3: Use the key value to partition the SAR image. The specific steps are as follows: Use the key value to get the threshold L of the low-light area of the brightness image T and the threshold L of the high illumination area of the brightness image H , and the pixel brightness in the SAR image is less than L T The pixel brightness is greater than L H The pixel brightness is between L T and L H Those in between are classified as medium illumination areas; Step 24: The partitioned SAR image is compressed using an adaptive logarithmic method to obtain the display brightness after tone mapping.
4. The self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets according to claim 3, characterized in that The logarithmic mean Y of the SAR image containing the ship target av It is expressed as: Among them, Y(x) represents the brightness channel of the image converted into XYZ space, W is the total number of pixels in the image, δ represents a positive number, and δ is 10 -4 , x represents the pixel position index in the image.
5. The self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets according to claim 4, characterized in that The key value is expressed as: Among them, Y max Indicates the maximum brightness value of the pixel, Y min Indicates the minimum brightness value of a pixel.
6. The self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets according to claim 5, characterized in that The L T and L H It is expressed as: L T =L max -(0.9+0.1key)(L max -L min ) L H =L min +[0.6+0.4×(1-key)](L max -L min ) Among them, L max and L min The images are normalized to the maximum and minimum values between 0 and 1, respectively.
7. The self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets according to claim 6, characterized in that The display brightness after tone mapping is expressed as: Among them, L i (x, y) represents the display brightness after tone mapping, L w (x,y) and L w,max They represent the original brightness and the maximum value of the original brightness respectively. q and k are constants. i ∈[1,∞],k i ∈[1,∞]; i=T,W,H; j=1,2, when j=1, S1=10, j=2, S2=2, S represents the base of the logarithmic mapping.
8. The self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets according to claim 1, characterized in that The clustering algorithm in step 3 is the K-means clustering algorithm, which is specifically expressed as follows: Among them, x i represents the i-th sample, that is, the i-th pixel in the image, and j represents x i The cluster, p j represents the center point corresponding to the cluster, and n represents the total number of samples.
9. The self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets according to claim 1, characterized in that The specific steps of step 4 are: Step 41: Use bilateral filtering algorithm to process the extracted target rectangular contour; Step 42: Cluster the bilaterally filtered image to obtain a refined target contour; Step 43: Map the refined target contour back to the original SAR image to obtain a mapped image.
10. The self-supervised deep learning sidelobe and clutter suppression method for SAR image ship targets according to claim 9, characterized in that The bilateral filtering algorithm is expressed as: Among them, w(i,j,k,l) represents the weight function in the bilateral filtering algorithm, d(i,j,k,l) represents the weighted combination of the values of the neighborhood pixels, r(i,j,k,l) represents the range template coefficient, (k,l) represents the center coordinates of the template window, (i,j) represents the coordinates of the other coefficients of the template window, σ d represents the standard deviation of the Gaussian function, f(i,j) represents the pixel value of the image at point (i,j), f(k,l) represents the pixel value of the image at point (k,l), σ r Represents the standard deviation of the Gaussian function.
Citation Information
Patent Citations
Image tone mapping method based on regional self-adaptive self-supervised learning
CN112085668A
Space environment high dynamic range imaging method based on multi-branch network
CN114998141A
Method, device and equipment for detecting surface defects of copper strip
CN118275449A
Radar image processing device, radar image processing method, and radar image processing program
US20200242738A1
