Ship wake detection method and system based on SAR image

Abstract

The invention relates to the technical field of SAR image detection and analysis, and provides a ship wake detection method and system based on an SAR image, and the method comprises the steps: 1, obtaining the SAR image; step 2, preprocessing the SAR image to obtain preprocessed SAR image data; step 3, performing multi-scale decomposition on the preprocessed SAR image data by adopting non-subsampled shearlet transform; step 4, performing GMC sparse regularization processing on the decomposed high-frequency sub-bands; 5, establishing a layered Bayesian inference model, and dynamically optimizing regularization parameters in the GMC; step 6, performing Radon transformation on the feature map # imgabs0 # after sparse optimization to extract wake linear features, and removing false detection and invalid features in combination with a verification step; and step 7, storing and visually displaying a final result. According to the method, the sparsity of the wake features can be enhanced, and the detection robustness is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of SAR image detection and analysis, and in particular to a ship wake detection method and system based on SAR images. Background Art

[0002] SAR (Synthetic Aperture Radar) image ship wake detection has important application value in the fields of ocean monitoring, illegal ship activity tracking and national defense security. Since SAR images can be imaged under all-weather and all-time conditions, they have irreplaceable advantages in complex marine environments. The formation of ship wakes includes turbulent wakes, narrow V wakes and Kelvin wakes, which not only reflect the motion state of the ship, but also reveal the speed, direction and type of the target. However, due to the inherent speckle noise in SAR images and the high complexity of the ocean background, the wake signal is usually weak and easily masked by background noise. How to achieve high-precision and robust wake detection in complex backgrounds and strong noise environments is the core challenge of current technology.

[0003] Existing SAR image ship wake detection technologies mainly focus on image decomposition, denoising and feature extraction. Early Radon transform and Hough transform methods detect wakes by enhancing linear features in images, but there are problems of false detection and missed detection under complex backgrounds and noise interference. In recent years, local Radon transform combined with low-rank sparse decomposition methods have improved the robustness of detection, but are still limited by the insufficient utilization of multi-scale features and reliance on manual parameter adjustment. In addition, although deep learning methods (such as CNN) can automatically extract features, they rely on a large amount of labeled data and have high computing resource requirements, making it difficult to meet the requirements of real-time and portability.

[0004] Taking the prior art as an example, patent CN115932854A proposes a detection method based on Kelvin wake features, which processes SAR images using the Kelvin wake model by obtaining the motion parameters of the ship. This method has a good detection effect on specific wake types, but has poor adaptability to complex backgrounds and other types of wakes (such as turbulent wakes and narrow V wakes). Another patent, CN116977176A, improves the resolution of SAR images by multi-scale feature fusion. Although this technology has advantages in image enhancement, it fails to specifically target the extraction and detection of wake features, especially in strong noise backgrounds. There are still significant deficiencies. Summary of the invention

[0005] The present invention mainly solves the problems of high false detection rate, missed detection and insufficient robustness in the prior art of ship wake detection in complex background and high noise environment. To this end, the present invention proposes a ship wake detection method and system based on SAR images, which uses non-subsampled shearlet transform (NSST) for multi-scale and multi-directional feature separation to effectively extract the wake signal from the complex background; by introducing generalized minimax concave (GMC) sparse regularization and combining L1 norm and Huber function, the sparsity of wake features is further enhanced, and the detection robustness is significantly improved; at the same time, the hierarchical Bayesian inference technology optimizes the dynamic adjustment of regularization parameters, avoiding the negative impact of fixed parameter settings on detection accuracy.

[0006] The present invention provides a ship wake detection method based on SAR images, comprising:

[0007] Step 1, acquiring SAR images;

[0008] Step 2, preprocessing the SAR image to obtain preprocessed SAR image data;

[0009] Step 3, the preprocessed SAR image data is subjected to multi-scale decomposition using non-subsampled shearlet transform (NSST);

[0010] Step 4, performing GMC sparse regularization processing on the decomposed high frequency subband;

[0011] Step 5, establish a hierarchical Bayesian inference model and dynamically optimize the regularization parameters in GMC;

[0012] Step 6: Optimize the sparse feature map Perform Radon transform to extract the linear features of the tail, and combine the verification step to remove false detection and invalid features;

[0013] Step 7: Store and visualize the final results.

[0014] Furthermore, in step 2, Wiener filtering is used to remove speckle noise in the SAR image. The formula is as follows:

[0015]

[0016] Among them, Y(x, y) represents the pixel value of the input SAR image, μ local is the local mean, is the local variance, is the noise variance.

[0017] Further, step 3 includes the following steps 301 to 305:

[0018] Step 301, input the preprocessed SAR image Y;

[0019] Step 302, obtaining the decomposition scale K and the direction number θ set by the user;

[0020] Step 303, using a non-subsampled Laplacian pyramid to preliminarily decompose the image to obtain a low-frequency subband and a high-frequency subband;

[0021] Step 304: high Perform directional decomposition and extract the wake features in different directions;

[0022] Step 305: store the low-frequency sub-band and the high-frequency sub-band obtained by decomposition.

[0023] Further, step 4 includes the following steps 401 to 404:

[0024] Step 401, setting sparse coefficients, weight matrix and regularization parameters, defining convergence threshold and maximum number of iterations;

[0025] Step 402, performing sparse optimization iterative update;

[0026] The objective function of the sparse optimization is:

[0027]

[0028] Where Y is the input high frequency subband; C is the sparse representation basis matrix; X is the sparse coefficient to be optimized; λ>0 is the regularization parameter; ψ(X) is the GMC regularization term;

[0029] The expression of the generalized minimax concave regularization is:

[0030] ψ(X)=||X|| 1 -S B (X),

[0031] in,

[0032]

[0033] Among them, S B (X) is an auxiliary term based on a concave function; v is an auxiliary variable; B is a weight matrix;

[0034] Step 403, when the sparse coefficients meet the following conditions, the iteration is stopped:

[0035] ||X (t+1) -X (t) || 2 <∈,

[0036] Among them, ∈>0 is the convergence threshold;

[0037] Step 404: The sparsely optimized high frequency subband X is output.

[0038] Furthermore, step 5 specifically includes the following steps 501 to 505:

[0039] Step 501, establishing a posterior probability model of the regularization parameter λ through hierarchical Bayesian inference;

[0040] The posterior probability formula is as follows:

[0041] p(X, λ∣Y)∝p(Y∣X)p(X∣λ)p(λ)

[0042] Among them, p(Y|X) is the likelihood function; p(X|λ) is the sparse prior; p(λ) is the prior distribution of the regularization parameter;

[0043] Step 502, defining a prior distribution and likelihood function;

[0044] The regularization parameter λ is assumed to follow the Gamma distribution, which is:

[0045]

[0046] Among them, α and β are hyperparameters that control the shape and scale of the distribution respectively; Γ(α) is the Gamma function;

[0047] The likelihood function p(Y|X) is assumed to be a Gaussian distribution and is of the form:

[0048]

[0049] Among them, σ 2 is the noise variance;

[0050] The sparse prior p(X|λ) is assumed to be Laplacian distributed:

[0051] p(X|λ)∝exp(-λ||X|| 1 )

[0052] Step 503, by maximizing the posterior probability, dynamically adjust the regularization parameter λ, and the update rule is as follows:

[0053]

[0054] Where N is the number of data points; ||X (t) || 1 is the current sparse coefficient L 1 norm;

[0055] Step 504: When the change of the regularization parameter is less than a preset threshold, the iteration is stopped:

[0056] |λ (t+1) -λ (t) |<∈

[0057] Where ∈ is the set convergence threshold, indicating that the update will stop when the parameter change is less than this value;

[0058] Step 505: output the optimized regularization parameters and the optimal solution.

[0059] Further, step 6 includes the following steps 601 to 607:

[0060] Step 601: Input is the sparse optimized feature map

[0061] Step 602, applying an angle range restriction to the sparsely optimized feature map, extracting features in the target direction, and generating a restricted feature map;

[0062] Step 603: Optimize the sparse feature map after restriction Apply Radon transform to map linear features in the image to the transform domain:

[0063]

[0064] Where δ(·) represents the Dirac δ function; r represents the distance between the linear feature and the origin; θ represents the direction angle of the linear feature;

[0065] Step 603, detecting significant peaks in the Radon transform domain, and extracting turbulent wakes and narrow V wakes;

[0066] Step 604, performing an inverse Radon transform on the detected feature points to remap the features back to the image space;

[0067] Step 605, performing multi-layer verification;

[0068] Step 606, classifying and labeling the trails according to the direction and geometric characteristics of the features;

[0069] Step 607, store the detection results, and output the classification, strength and direction information of the detected trails.

[0070] Correspondingly, the present invention also provides a ship wake detection system based on SAR images, comprising: a data acquisition module, a preprocessing module, an NSST decomposition module, a sparse optimization module, a hierarchical Bayesian inference module, a wake detection module and a data output module;

[0071] The data acquisition module is used to acquire SAR images;

[0072] The preprocessing module is used to preprocess the SAR image to obtain preprocessed SAR image data;

[0073] The NSST decomposition module is used to perform multi-scale decomposition on the pre-processed SAR image data using non-subsampled shearlet transform;

[0074] The sparse optimization module is used to perform GMC sparse regularization processing on the decomposed high frequency sub-bands;

[0075] The hierarchical Bayesian inference module is used to establish a hierarchical Bayesian inference model and dynamically optimize the regularization parameters in the GMC;

[0076] The tail detection module is used to detect the feature map after sparse optimization. Perform Radon transform to extract linear features, and combine with verification steps to remove false positives and invalid features;

[0077] The data output module is used to store and visually display the final results.

[0078] The present invention provides a ship wake detection method and system based on SAR images, which achieves higher accuracy, strong robustness and automation capability that are difficult to achieve with the existing technology by adopting non-subsampled shearlet transform (NSST), generalized minimax concave (GMC) sparse regularization and hierarchical Bayesian inference technology. Compared with the existing technology, it has the following advantages:

[0079] 1. Multi-scale decomposition and direction enhancement to improve the accuracy of wake detection.

[0080] This application uses NSST to perform multi-scale and multi-directional decomposition of SAR images, extract feature information of different scales and directions, effectively separate background noise and tail features at different scales, enhance the directional expression of tail signals, and make tail signals stand out more clearly from background noise, thereby improving the detection accuracy in complex backgrounds. Compared with the traditional Radon transform method, NSST reduces noise interference while retaining the details of tail features.

[0081] 2. Sparse optimization and feature enhancement to improve the robustness of weak wake signals.

[0082] The high-frequency subband is processed through GMC regularization, and the sparse characteristics are used to enhance the significance of the tail signal and suppress background noise. GMC sparse regularization is combined with hierarchical Bayesian inference to dynamically adjust the regularization parameters to improve the adaptability of the model in complex backgrounds and high-noise environments. Compared with traditional L1 regularization, the GMC regularization of this application achieves a stronger suppression ability for weak tail features by combining the Huber function, so that the detection is still highly robust in an environment with strong noise interference, avoiding the performance bottleneck caused by fixed parameters in traditional methods.

[0083] 3. Adaptive optimization of parameters to reduce reliance on manual adjustments.

[0084] The hierarchical Bayesian inference module dynamically optimizes the regularization parameters through the Gamma distribution, automatically adjusts the regularization weights of the model, and avoids the degradation of detection performance caused by improper parameter selection. Compared with the manual adjustment of parameters in the existing methods, this application significantly improves the automation level of the detection process.

[0085] 4. Automate the testing process to improve work efficiency.

[0086] This application has built a fully automated workflow from data preprocessing, wake detection to result output, built a complete automated detection system, reduced the steps of manual intervention, and greatly improved detection efficiency and accuracy. Experimental results show that the detection method of the present invention has good applicability in real-time detection scenarios and can meet the needs of rapid detection in complex marine environments.

[0087] 5. Widely applicable scenarios, enhancing the adaptability of the system.

[0088] The technical solution of the present application shows good adaptability to different types of wakes (such as turbulent wakes, narrow V wakes and Kelvin wakes), and can process a variety of SAR image formats and data sources, and has wide practicality.

[0089] 6. The reliability of the results is improved, which facilitates subsequent analysis.

[0090] The detection results are verified by combining the trail detection module with Radon transform and inverse transform, effectively reducing false detection and missed detection. The result output module supports the visualization of the detection results and storage in multiple formats, which is convenient for users to conduct subsequent analysis.

[0091] In summary, the technical solution of the present application effectively solves the problems of low detection accuracy, poor robustness and parameter dependence on manual adjustment in complex backgrounds and high noise environments in the existing technology by adopting innovative technical means such as NSST, GMC sparse regularization and hierarchical Bayesian inference, and significantly improves the performance and practicality of the detection system. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 It is a flow chart of the implementation of the ship wake detection method based on SAR images provided by the present invention;

[0093] Figure 2 It is a schematic diagram of NSST decomposition provided by the present invention;

[0094] Figure 3 is an example of a raw SAR image;

[0095] Figure 4 This is an example of a high-frequency component image after NSST decomposition;

[0096] Figure 5 This is an example of a low-frequency component image after NSST decomposition;

[0097] Figure 6 This is an example of the directional component image after NSST decomposition;

[0098] Figure 7 is an example of a wake detection result;

[0099] Figure 8 This is an example of the Radon domain graph corresponding to the detection result. DETAILED DESCRIPTION

[0100] In order to make the technical problems solved by the present invention, the technical solutions adopted and the technical effects achieved clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention are shown in the accompanying drawings, rather than all the contents.

[0101] Embodiment 1

[0102] like Figure 1 As shown, a ship wake detection method based on SAR images provided by an embodiment of the present invention includes the following process:

[0103] Step 1: Acquire SAR images.

[0104] Obtain SAR images through satellite ground stations or local storage devices. The original image is as follows Figure 3 shown.

[0105] Step 2: preprocess the SAR image to obtain preprocessed SAR image data.

[0106] The SAR image is preprocessed to implement filtering and denoising, remove speckle noise, and obtain preprocessed SAR image data. Specifically, Wiener filtering is used to remove speckle noise in SAR images. The formula is as follows:

[0107]

[0108] Among them, Y(x, y) represents the pixel value of the input SAR image, μ local is the local mean, is the local variance, is the noise variance. The local variance here indicates the degree of variation of pixel values ​​in a local area of ​​the image, which is used to describe the noise characteristics of the area. By estimating the local area, the intensity of noise removal can be adjusted more accurately to avoid over-smoothing that affects image details.

[0109] Step 3: The preprocessed SAR image data is subjected to multi-scale decomposition using non-subsampled shearlet transform (NSST).

[0110] In this step, NSST can effectively extract the tail features and distinguish them from the background information by decomposing the image into high-frequency sub-bands and low-frequency sub-bands. high，k，θ Contains tail features, low frequency subband Y low The background information of the image is retained to provide support for subsequent sparse optimization and trail detection.

[0111] Nonsubsampled Shearlet Transform (NSST) is an advanced multi-scale, multi-directional image decomposition method that can effectively preserve the edge and detail features in SAR images while suppressing background noise. Figure 2 As shown, step 3 includes the following steps 301 to 305:

[0112] Step 301: input the preprocessed SAR image Y.

[0113] First, the SAR image Y after preprocessing such as noise removal is input into the system to prepare for NSST decomposition.

[0114] Step 302: Obtain the decomposition scale K and the direction number θ set by the user.

[0115] The decomposition scale K determines the number of decomposition levels, and is commonly selected as K=3 or K=4, which is adjusted according to the resolution of the image;

[0116] The direction number θ is usually set to θ=8, that is, the image is decomposed into 8 main directions.

[0117] These parameters determine the decomposition fineness of NSST and the range of directional feature extraction, adapting to different image features.

[0118] Step 303: Use a non-subsampled Laplacian pyramid (NSLP) to preliminarily decompose the image to obtain a low-frequency sub-band and a high-frequency sub-band.

[0119] NSLP is a decomposition method in NSST, which aims to decompose the image into sub-bands of multiple scales. The low-frequency sub-bands preserve the background information, while the high-frequency sub-bands retain the detailed features of the image. In the NSST framework, NSLP is used as a preliminary decomposition step to provide a basis for subsequent multi-scale and multi-directional processing. By decomposing the image by NSLP, the low-frequency sub-band Y low and high frequency sub-band Y high :

[0120] Y low , Y high =NSLP(Y)

[0121] Among them, Y low Contains the global background information of the image; Y high Contains details and edge features in the image, especially trail features.

[0122] This step removes unnecessary low-frequency information and retains the high-frequency characteristics of the tail to facilitate subsequent processing.

[0123] Step 304: high Perform directional decomposition and extract the trail features in different directions.

[0124] In the high frequency subband, the directional filter G in NSST is used θ (i, j) is decomposed in multiple directions to further extract the directional characteristics of the wake. The directional decomposition formula is as follows:

[0125]

[0126] Among them, G θ (i, j) is the shear wave basis function, which determines the direction of decomposition; θ is the number of directions, usually set to 8, indicating that decomposition is performed in 8 main directions; Y k (x+i, y+j) is the high frequency subband of the kth scale.

[0127] Directional filter G θ The high frequency subband Y high Projection is performed in different directions to obtain the tail features in each direction. This step makes the tail features more clear and directional, such as Figure 6 shown.

[0128] Step 305: store the low-frequency sub-band and the high-frequency sub-band obtained by decomposition.

[0129] Among them, Y low Store and use for background information suppression; Yhigh，k，θ Stored and used for enhancement of wake features.

[0130] These decomposed subbands will be used in subsequent sparse optimization to process the background and tail respectively. high，k，θ (like Figure 4 as shown) and low frequency sub-band Y low (like Figure 5 As shown in Figure 3, the sparse optimization algorithm is provided to the subsequent sparse optimization algorithm for further analysis and tail detection.

[0131] Step 4: Perform GMC sparse regularization processing on the decomposed high-frequency sub-band.

[0132] GMC sparse regularization processing is the core process of the present invention. It uses sparse expression to enhance the tail features in SAR images and suppress background noise, which can enhance the sparsity of the tail signal. Through generalized minimax concave (GMC) regularization, the feature enhancement of the high-frequency sub-band is further achieved and the background noise is suppressed, ensuring the robustness and accuracy of the results. In this process, the present invention constructs an inverse problem model to express the imaging process of SAR images. Specifically, the input high-frequency sub-band data Y can be expressed as the following inverse problem formula:

[0133] Y=CX+N

[0134] Where Y is the high frequency subband of the observed SAR image, C is the sparse representation basis matrix, X is the sparse coefficient to be optimized, and N is the noise term. This formula represents the linear mapping of the sparse coefficient X through the sparse representation basis matrix C and the addition of noise N.

[0135] Step 4 includes the following steps 401 to 404:

[0136] Step 401, setting the sparse coefficient, weight matrix and regularization parameter, and defining the convergence threshold and the maximum number of iterations.

[0137] Specifically, set the sparse coefficient X (0) =0, weight matrix B (0) =I and regularization parameter λ>0; define the convergence threshold ∈ and the maximum number of iterations T.

[0138] Step 402, perform sparse optimization iterative update.

[0139] The objective function of the sparse optimization is:

[0140]

[0141] Where Y is the input high-frequency subband; C is the sparse representation basis matrix; X is the sparse coefficient to be optimized; λ>0 is the regularization parameter; ψ(X) is the GMC regularization term used to enhance sparsity.

[0142] The expression of the generalized minimax concave regularization is:

[0143] ψ(X)=||X|| 1 -S B (X),

[0144] in,

[0145]

[0146] Among them, S B (X) is an auxiliary term based on a concave function; v is an auxiliary variable; B is a weight matrix that controls the balance between sparsity and smoothness. It can effectively enhance the sparsity of the trail feature and suppress background interference. Step 402 specifically includes the following steps 4021 to 4024:

[0147] Step 4021, smoothing items Perform a gradient descent update:

[0148]

[0149] in:

[0150]

[0151] Among them, μ>0 is the gradient descent step size; X (t+1 / 2) is the intermediate result.

[0152] Step 4022, apply a soft threshold operation to the intermediate result to achieve L1 sparsification:

[0153]

[0154] The soft threshold function is defined as:

[0155] soft(z, λ)=sign(z)·max(|z|-λ, 0)

[0156] The present invention uses a forward-backward splitting algorithm (FB algorithm) to iteratively solve, and each update is completed by gradient descent and soft threshold operation.

[0157] Step 4023, concave portion S B (X) Introduce auxiliary variable v and decompose it into sub-problems:

[0158]

[0159] The problem is solved by the fast projection algorithm.

[0160] Step 4024, after each iteration, dynamically adjust the weight matrix B according to the optimization result:

[0161]

[0162] Here ∈>0 is used to avoid the denominator being zero.

[0163] Step 403, when the sparse coefficients meet the following conditions, the iteration is stopped:

[0164] ‖X (t+1) -X (t) || 2 <∈,

[0165] Where ∈>0 is the convergence threshold.

[0166] Step 404: The sparsely optimized high frequency subband X is outputted for subsequent tail detection.

[0167] In this process, X represents the high-frequency subband coefficient obtained through sparse optimization, and in the subsequent step 5, we will further optimize the coefficient to obtain the final inference result

[0168] Step 5: Establish a hierarchical Bayesian inference model and dynamically optimize the regularization parameters in GMC.

[0169] It should be noted that in this step, sparse optimization (step 4) and Bayesian inference (step 5) are performed alternately. Specifically, first in the sparse optimization phase of step 4, the current regularization parameter λ is used (t) Optimize the sparse coefficient X to obtain the sparse solution X (t) At this point, the regularization parameter λ plays a role in controlling sparsity. A larger λ value will cause more coefficients to approach zero, resulting in a sparser solution. Next, in step 5, the hierarchical Bayesian inference phase, based on the current sparse coefficients X (t) , the present invention dynamically adjusts the regularization parameter λ by maximizing the posterior probability, so that it adapts to the noise level and tail characteristics in the image. Finally, these two processes are performed alternately, and the updated λ is brought into the next round of sparse optimization to further optimize the sparse coefficient X. Each sparse optimization provides a more accurate X for Bayesian inference, and each Bayesian inference makes λ more adaptable to the characteristics of the current data.

[0170] The core goal of this step is to dynamically adjust the regularization parameter λ through hierarchical Bayesian inference, thereby achieving adaptive optimization and obtaining a final optimal solution. That is, the most appropriate sparse coefficient is obtained by maximum a posteriori inference. Especially in the noise changes and complex backgrounds in SAR image processing, the accuracy and robustness of the optimization process are ensured. Step 5 specifically includes the following steps 501 to 505:

[0171] Step 501, establish a posterior probability model of the regularization parameter λ through hierarchical Bayesian inference.

[0172] The posterior probability model of the regularization parameter λ is established, and the goal is to estimate the most appropriate λ by maximizing the posterior probability.

[0173] The posterior probability formula is as follows:

[0174] p(X, λ∣Y)∝p(Y∣X)p(X∣λ)p(λ)

[0175] Among them, p(Y|X) is the likelihood function, which represents the relationship between the observed data and the model; p(X|λ) is the sparse prior, which represents the sparsity of the data; p(λ) is the prior distribution of the regularization parameter, and it is usually assumed that λ follows the Gamma distribution.

[0176] Step 502, define the prior distribution and likelihood function.

[0177] The regularization parameter λ is assumed to follow the Gamma distribution, which is:

[0178]

[0179] Among them, α and β are hyperparameters that control the shape and scale of the distribution respectively; Γ(α) is the Gamma function.

[0180] The likelihood function p(Y|X) is assumed to be a Gaussian distribution and is of the form:

[0181]

[0182] where σ 2 is the noise variance.

[0183] The sparse prior p(X|λ) is assumed to be Laplacian distributed, indicating the sparsity of the data:

[0184] p(X|λ)∝exp(-λ||X|| 1 )

[0185] Step 503, by maximizing the posterior probability, dynamically adjust the regularization parameter λ, and the update rule is as follows:

[0186]

[0187] Where N is the number of data points; ||X (t) || 1 is the current sparse coefficient L 1 Norm.

[0188] Through this formula, the regularization parameter λ is updated in each iteration according to the sparsity of the data, thereby adaptively adjusting the regularization strength during the optimization process.

[0189] Step 504: When the change of the regularization parameter is less than a preset threshold, the iteration is stopped:

[0190] |λ (t+1) -λ (t) |<∈

[0191] Where ∈ is the set convergence threshold, indicating that the update is stopped when the parameter change is less than this value.

[0192] Step 505: output the optimized regularization parameters and the optimal solution.

[0193] Output the dynamically optimized regularization parameter λ as the input for the next step of optimization. By maximizing the posterior probability, the currently updated regularization parameter λ is used. (t+1) , optimize the sparse coefficient X, and finally obtain the optimal sparse coefficient solution (The sparsely optimized feature map is a feature set that has been processed by sparse optimization, focusing on the enhancement of the tail signal).

[0194] Step 6: Optimize the sparse feature map Radon transform is performed to extract the linear features of the tail, and verification steps are combined to remove false detections and invalid features.

[0195] The main task of this step is to extract and verify the linear features of the feature map after sparse optimization and feature enhancement. The linear features are mapped to the transform domain through Radon transform, and combined with inverse transform and multi-layer verification mechanism to ensure the accuracy and robustness of the detection results. Step 6 includes the following steps 601 to 607:

[0196] Step 601: Input is the sparse optimized feature map

[0197] In this step, the input sparse optimized feature map This figure contains the tail features optimized by GMC regularization. is a feature set obtained after sparse optimization processing, focusing on the enhancement of the wake signal. It has undergone multi-scale decomposition and sparse optimization and can accurately reflect the trail features in the image.

[0198] Step 602, impose an angle range restriction on the sparsely optimized feature map, extract features in the target direction, and generate a restricted feature map.

[0199] In this step, in order to reduce noise interference and focus on the direction of the target trail characteristics, Apply angle range restrictions, extract features in the target direction, and generate restricted feature maps

[0200]

[0201] Among them, θ range =[-20°, 20°] ° The characteristics of the target direction are retained and background noise is filtered. By limiting the angle, the trail characteristics in a specific direction can be effectively focused and the influence of irrelevant information can be reduced.

[0202] Step 603: Optimize the sparse feature map after restriction Apply Radon transform to map linear features in the image to the transform domain:

[0203]

[0204] Among them, δ(·) represents the Dirac δ function; r represents the distance between the linear feature and the origin; θ represents the direction angle of the linear feature. Through this transformation, the linear features in the image (such as tails) will appear as obvious peaks in the Radon domain, which is convenient for subsequent feature extraction and analysis.

[0205] Step 603: detect significant peaks in the Radon transform domain and extract turbulent wakes and narrow V wakes.

[0206] The turbulent wake: extract the main wake based on its intensity and geometric characteristics.

[0207] The narrow V-wake: searching for a second wake within a range of ±4° near the turbulent wake; and determining whether the angle between the two narrow V-wakes is Δθ≤8°.

[0208] The Kelvin wake detection: searching for Kelvin wakes in the region of 10°≤|θ|≤20° around the turbulent wake; according to the intensity threshold T Kelvin Extract the peak of the Kelvin tail.

[0209] Step 604: perform inverse Radon transform on the detected feature points to remap the features back to the image space.

[0210] After extracting the target features in the Radon transform domain, we remap these features back to the image space through the inverse Radon transform to obtain the reconstructed trail feature map:

[0211] Y′(x, y)=f 0 πR(r,θ)δ(r-xcosθ-ysinθ)dθ

[0212] Step 605, perform multi-layer verification.

[0213] Orientation consistency check: Ensures that the feature orientation matches the expected pattern of vessel motion.

[0214] Geometric morphology analysis: Determine whether the distribution of characteristic points conforms to the geometric characteristics of turbulent wakes or Kelvin wakes.

[0215] Intensity and length matching: Filter out low-intensity pseudo features or features that do not meet the trail length requirements.

[0216] Step 606, classify and label the trails according to the direction and geometric characteristics of the features.

[0217] The wakes are classified into turbulent wakes, narrow V wakes and Kelvin wakes.

[0218] Step 607, store the detection results, and output the classification, strength and direction information of the detected trails.

[0219] Stores detection results and supports 2D and 3D visualization. Outputs the classification, intensity, and direction information of detection trails.

[0220] Step 7: Store and visualize the final results.

[0221] This step is to store and display the detection results in a form that is easy to understand and analyze. It supports multiple output formats and visualization methods to provide users with efficient presentation of wake detection results.

[0222] Step 701, receiving a detection result, wherein the detection result includes: trail feature data and detected image data.

[0223] The trail characteristic data includes parameters such as position, direction, and length.

[0224] The detected image data: the trail feature enhanced image after sparse optimization.

[0225] Step 702: Generate data output in multiple formats according to user requirements.

[0226] GeoTIFF format: contains geocoded images of wake detection results, which is convenient for integration with GIS systems;

[0227] CSV format: records the tabular data of wake parameters for statistical analysis;

[0228] JSON format: structured test results, used for API calls or other secondary development needs.

[0229] Step 703: Visualize the detection results.

[0230] 2D visualization: Displays the detection results superimposed on the original SAR image; annotates the main features of the trail (such as direction, position).

[0231] 3D visualization: Build a 3D model of the wake features for dynamic observation and analysis.

[0232] The method of the present invention utilizes sparse expression to enhance the tail features in SAR images while suppressing background noise. The feature enhancement of high-frequency subbands is achieved through generalized minimax concave (GMC) regularization, and the sparse coefficients are dynamically adjusted in combination with an iterative optimization algorithm to ensure the robustness and accuracy of the results. By combining non-subsampled shearlet transform (NSST) and generalized minimax concave (GMC) sparse regularization, and introducing hierarchical Bayesian inference, multi-scale enhancement and dynamic optimization adjustment of tail signals are achieved. The method has the characteristics of high detection accuracy, strong robustness and high degree of automation.

[0233] This example demonstrates a complete implementation system based on this patented technology, which is applied to SAR image ship wake detection, with the goal of achieving high-precision detection under complex background and high noise conditions. The data source is C-band SAR satellite imagery with a resolution of 3 meters and a data format of GeoTIFF, including turbulent wakes, narrow V wakes, and Kelvin wakes (such as Figure 7 and 8 shown).

[0234] Experimental results show that the detection accuracy and robustness of the proposed method for weak wake signals in complex backgrounds are significantly better than existing methods, providing a more efficient and reliable solution for ship wake detection in SAR images.

[0235] Embodiment 2

[0236] An embodiment of the present invention provides a ship wake detection system based on SAR images. The ship wake detection system based on SAR images in this embodiment corresponds to the ship wake detection method based on SAR images in embodiment one, and can execute the ship wake detection method based on SAR images in embodiment one, which specifically includes a data acquisition module, a preprocessing module, an NSST decomposition module, a sparse optimization module, a hierarchical Bayesian inference module, a wake detection module and a data output module.

[0237] The data acquisition module is used to acquire SAR images. The data acquisition module acquires SAR image data from a satellite ground station or a local storage device, provides multi-source input, and ensures that the system can process different types of SAR image data.

[0238] The preprocessing module is used to preprocess the SAR image to obtain preprocessed SAR image data. The preprocessing module uses multi-core parallel processing technology to ensure the real-time data processing. After removing speckle noise through Wiener filtering, the filtered image data will be transmitted to the NSST decomposition module for subsequent processing.

[0239] The NSST decomposition module is used to perform multi-scale decomposition on the pre-processed SAR image data using non-subsampled shearlet transform. The NSST decomposition module performs multi-scale decomposition on the pre-processed SAR image through non-subsampled shearlet transform (NSST) to generate low-frequency sub-bands and high-frequency sub-bands at different scales and directions, so as to effectively extract the tail features.

[0240] The sparse optimization module is used to perform GMC sparse regularization processing on the decomposed high-frequency sub-bands. The sparse optimization module performs generalized minimax concave (GMC) sparse regularization processing on the decomposed high-frequency sub-bands, and enhances the sparsity of the tail feature through gradient descent and soft threshold operations to obtain the feature map of the optimal solution and suppress background noise.

[0241] The hierarchical Bayesian inference module is used to establish a hierarchical Bayesian inference model and dynamically optimize the regularization parameter in the GMC. The hierarchical Bayesian inference module dynamically optimizes the regularization parameter λ by maximizing the posterior probability according to the current sparsity coefficient X, thereby further improving the accuracy and robustness of the sparse optimization.

[0242] The tail detection module is used to detect the feature map after sparse optimization. The Radon transform is used to extract linear features, and the verification step is combined to remove false positives and invalid features. The trail detection module applies the Radon transform to extract linear features from the sparsely optimized feature map, and removes false positives and invalid features through the inverse Radon transform and multi-layer verification mechanism to ensure the accuracy and robustness of the detection results.

[0243] The data output module is used to store and visualize the final results. The data output module stores the final detection results and outputs them in GeoTIFF, CSV, JSON and other formats, supports 2D and 3D visualization, and provides information such as the classification, intensity, and direction of the trails.

[0244] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that modifying the technical solutions described in the aforementioned embodiments, or replacing some or all of the technical features therein by equivalents, does not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A ship wake detection method based on SAR images, characterized in that: include: Step 1, acquiring SAR images; Step 2, preprocessing the SAR image to obtain preprocessed SAR image data; Step 3, using non-subsampled shearlet transform to perform multi-scale decomposition on the pre-processed SAR image data; Step 4, performing GMC sparse regularization processing on the decomposed high frequency subband; Step 5, establish a hierarchical Bayesian inference model and dynamically optimize the regularization parameters in GMC; Step 6: Optimize the sparse feature map Perform Radon transform to extract the linear features of the tail, and combine the verification step to remove false detection and invalid features; Step 7: Store and visualize the final results.

2. The ship wake detection method based on SAR images according to claim 1 is characterized in that: In step 2, Wiener filtering is used to remove speckle noise in the SAR image. The formula is as follows: Among them, Y(x, y) represents the input SAR image pixel value, μ local is the local mean, is the local variance, is the noise variance.

3. The ship wake detection method based on SAR images according to claim 1 is characterized in that: Step 3 includes the following steps 301 to 305: Step 301, input the preprocessed SAR image Y; Step 302, obtaining the decomposition scale K and the direction number θ set by the user; Step 303, using a non-subsampled Laplacian pyramid to preliminarily decompose the image to obtain a low-frequency subband and a high-frequency subband; Step 304: high Perform directional decomposition and extract the wake features in different directions; Step 305: store the low-frequency sub-band and the high-frequency sub-band obtained by decomposition.

4. The ship wake detection method based on SAR images according to claim 3 is characterized in that: Step 4 includes the following steps 401 to 404: Step 401, setting sparse coefficients, weight matrix and regularization parameters, defining convergence threshold and maximum number of iterations; Step 402, performing sparse optimization iterative update; The objective function of the sparse optimization is: Where Y is the input high frequency subband; C is the sparse representation basis matrix; X is the sparse coefficient to be optimized; λ>0 is the regularization parameter; ψ(X) is the GMC regularization term; The expression of the generalized minimax concave regularization is: ψ(X)=||X||1-S B (X), in, Among them, S B (X) is an auxiliary term based on a concave function; v is an auxiliary variable; B is a weight matrix; Step 403, when the sparse coefficients meet the following conditions, the iteration is stopped: ||X (t+1) -X (t) ||2<∈, Among them, ∈>0 is the convergence threshold; Step 404: The sparsely optimized high frequency subband X is output.

5. The ship wake detection method based on SAR images according to claim 4 is characterized in that: Step 5 specifically includes the following steps 501 to 505: Step 501, establishing a posterior probability model of the regularization parameter λ through hierarchical Bayesian inference; The posterior probability formula is as follows: p(X, λ∣Y)∝p(Y∣X)p(X∣λ)p(λ) Among them, p(Y|X) is the likelihood function; p(X|λ) is the sparse prior; p(λ) is the prior distribution of the regularization parameter; Step 502, defining a prior distribution and likelihood function; The regularization parameter λ is assumed to follow the Gamma distribution, which is: Among them, α and β are hyperparameters that control the shape and scale of the distribution respectively; Γ(α) is the Gamma function; The likelihood function p(Y|X) is assumed to be a Gaussian distribution and is of the form: Among them, σ 2 is the noise variance; The sparse prior p(X|λ) is assumed to be Laplacian distributed: p(X|λ)∝exp(-λ||X||1) Step 503, by maximizing the posterior probability, dynamically adjust the regularization parameter λ, and the update rule is as follows: Where N is the number of data points; ||X (t) ||1 is the L1 norm of the current sparse coefficient; Step 504: When the change of the regularization parameter is less than a preset threshold, the iteration is stopped: |l (t+1) -l (t) |<∈ Where ∈ is the set convergence threshold, indicating that the update will stop when the parameter change is less than this value; Step 505: output the optimized regularization parameters and the optimal solution.

6. The ship wake detection method based on SAR images according to claim 5 is characterized in that: Step 6 includes the following steps 601 to 607: Step 601: Input is the sparse optimized feature map Step 602, applying an angle range restriction to the sparsely optimized feature map, extracting features in the target direction, and generating a restricted feature map; Step 603: Optimize the sparse feature map after restriction Apply Radon transform to map linear features in the image to the transform domain: Where δ(·) represents the Dirac δ function; r represents the distance between the linear feature and the origin; θ represents the direction angle of the linear feature; Step 603, detecting significant peaks in the Radon transform domain, and extracting turbulent wakes and narrow V wakes; Step 604, performing an inverse Radon transform on the detected feature points to remap the features back to the image space; Step 605, performing multi-layer verification; Step 606, classifying and labeling the trails according to the direction and geometric characteristics of the features; Step 607, store the detection results, and output the classification, strength and direction information of the detected trails.

7. A ship wake detection system based on SAR images, characterized in that: include: Data acquisition module, preprocessing module, NSST decomposition module, sparse optimization module, hierarchical Bayesian inference module, wake detection module and data output module; The data acquisition module is used to acquire SAR images; The preprocessing module is used to preprocess the SAR image to obtain preprocessed SAR image data; The NSST decomposition module is used to perform multi-scale decomposition on the pre-processed SAR image data using non-subsampled shearlet transform; The sparse optimization module is used to perform GMC sparse regularization processing on the decomposed high frequency sub-bands; The hierarchical Bayesian inference module is used to establish a hierarchical Bayesian inference model and dynamically optimize the regularization parameters in the GMC; The tail detection module is used to detect the feature map after sparse optimization. Perform Radon transform to extract linear features, and combine with verification steps to remove false positives and invalid features; The data output module is used to store and visually display the final results.

Images