Polarization SAR ship target detection method and device based on multi-channel network
By constructing a multi-channel polarimetric SAR ship target detection system, and utilizing the multi-channel characteristics of polarimetric SAR data and an improved YOLOv8 network structure, the problem of insufficient ship detection accuracy under complex sea conditions is solved, and high-precision ship target detection is achieved.
Patent Information
- Application Number
- CN202510925522.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-05
- Publication Date
- 2025-10-17
AI Technical Summary
Traditional ship detection methods based on polarization characteristics have limited accuracy in complex sea conditions, especially for small civilian vessels, which have a high rate of missed detection and cannot meet the needs of maritime supervision and search and rescue.
A polarimetric SAR ship target detection system based on a multi-channel network is constructed. By fusing features from the main and auxiliary channels, the system utilizes the multi-channel features of polarimetric SAR data, employs an improved YOLOv8 network structure for target detection, and combines an FPN structure for multi-scale feature extraction and detection output.
It improves the detection accuracy of civilian ships to 92.3% under complex sea conditions, effectively solving the problem of missed detection caused by small targets and clutter interference, and significantly improving the detection effect.
Smart Images
Figure CN120807887A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of computer neural networks and signal processing, in particular to a polarimetric SAR ship target detection method and device based on a multi-channel network. BACKGROUND
[0002] Polarimetric synthetic aperture radar (polarimetric SAR) can obtain the scattering matrix information of ground targets by transmitting and receiving electromagnetic waves of different polarization modes, and has a significant advantage in ship target detection. However, the detection accuracy of traditional ship detection methods based on polarization features (such as scattering mechanism decomposition, polarization covariance matrix statistical characteristics, etc.) is often limited in complex sea conditions (such as high sea conditions, strong clutter interference) or near-shore multi-target scenes. In particular, in the detection of fishing boats, yachts and other civilian ships, due to the small physical size of the target and the non-metallic composite material of the target, the polarization scattering characteristics of the target are often similar to those of sea clutter, resulting in a high miss rate and making it difficult to meet the needs of maritime supervision, maritime search and rescue and other civilian needs.
[0003] In recent years, deep learning has shown great potential in polarimetric SAR image interpretation, but deep learning has certain limitations. Most existing researches focus on modeling polarimetric SAR data as RGB optical image data, and fail to fully utilize the scattering characteristic information of the polarization channel. SUMMARY
[0004] The present application provides a polarimetric SAR ship target detection method and device based on a multi-channel network, which solves the technical problem of high miss rate and limited detection accuracy in complex sea conditions in the prior art, breaks through the bottleneck of the prior art, and has important significance for realizing high-precision detection in complex scenes.
[0005] In order to achieve the above-mentioned purpose, the technical scheme of the present application is as follows:
[0006] The present application provides a polarimetric SAR ship target detection method based on a multi-channel network, comprising the following steps:
[0007] S1. Constructing a polarimetric SAR ship target detection system based on a multi-channel network;
[0008] S2. Constructing a data set for training a polarimetric SAR ship target detection system based on a multi-channel network;
[0009] S3. Using the training set to optimize the weights of the polarimetric SAR ship target detection system based on the multi-channel network, and saving the trained weight parameters;
[0010] S4. Loading the pre-stored weight parameters based on the polarimetric SAR ship target detection system based on the multi-channel network to obtain the final detection system;
[0011] S5. The final detection system detects the polarized SAR ship target input and outputs the result.
[0012] Further, the polarized SAR ship target detection system in S1 comprises:
[0013] The preprocessing module annotates the original data set, divides it into a training set and a test set, and randomly rearranges it to improve the adaptability and generalization performance of the network.
[0014] The feature fusion module integrates different feature information, and the main channel and the auxiliary channel respectively pass through independent convolution blocks for shallow feature extraction. The output feature map is spliced in the channel dimension to generate fused high-order features.
[0015] The backbone feature extraction module extracts edge and texture bottom layer features through shallow convolution, and the deep residual module captures global semantic information of the target.
[0016] The pyramid feature extraction module uses the FPN structure to fuse the multi-scale features of the backbone network from top to bottom, combining high-level semantic information with low-level positioning accuracy.
[0017] The detection output module outputs the class probability of each anchor box through the convolution layer, predicts the offset of the bounding box, and evaluates the probability of the target existing in the predicted box.
[0018] Further, S2 comprises the following steps:
[0019] S21. Data preprocessing: extracting the main diagonal elements T11, T22, T33 of the coherence matrix of polarized SAR data as the main channel data input, and performing eight-component decomposition on the coherence matrix to obtain the auxiliary channel data input.
[0020] S22. Constructing a data set: annotating the target bounding box and class label, and dividing the training set and the test set.
[0021] Further, the main channel data input in S21 is the coherence matrix main diagonal element T11, T22, T33 component; the auxiliary channel data input is the surface scattering, volume scattering, even scattering, helical body scattering, rotation dihedral scattering, rotational dipole scattering, rotational 1 / 4 wave device scattering and mixed dipole scattering features obtained by eight-component decomposition.
[0022] Further, S3 comprises the following steps:
[0023] S31. Initialize the weight parameters: the network weight parameters of the feature fusion module, the backbone feature extraction module, the pyramid feature extraction module and the detection output module are initialized using a normal distribution with a mean of 0 and a variance of 0.01.
[0024] S32. Setting the training parameters of the polarization SAR ship target detection system based on the multi-channel network: adopting AdamW as the model training optimizer, setting the hyperparameters β1 of the model training optimizer to 0.95, β2 to 0.9, and the weight decay to 5x10 -4 The batch size is 16, and the maximum training step is 200;
[0025] S33. Calculate the loss function value Loss, which includes the coordinate prediction loss function, the confidence loss function, and the class prediction loss function.
[0026] S34. The termination condition for determining whether the training is finished is that the loss function value Loss ≤ 0.005 or 200 training rounds are reached.
[0027] The application also provides a polarization SAR ship target detection device based on a multi-channel network, comprising:
[0028] a memory for storing pre-processing data sets and network weight parameters; and
[0029] a processor configured to perform data preprocessing, dual-channel feature fusion of main channels and auxiliary channels, multi-scale feature extraction, and target detection output; and
[0030] an output interface for transmitting the detection results of target categories, bounding box coordinates, and confidence.
[0031] Further, the processor architecture of the polarization SAR ship target detection device based on the multi-channel network adopts an improved YOLOv8 network structure, which expands the single input channel into a dual-channel structure.
[0032] By adopting the above technical solutions, the application has the following advantages:
[0033] The application provides a polarization SAR ship target detection method and device based on a multi-channel network. By inputting the T11, T22, and T33 data of the polarization SAR data as the main channel data and the data after eight-component decomposition of the polarization SAR data as the auxiliary channel data, a multi-channel deep learning network is constructed, which can cooperatively mine the spatial-scattering correlation features of the polarization data and improve the detection capability of the model for ship targets. The detection accuracy for civilian ships under complex sea conditions reaches 92.3%, effectively solving the problem of missed detection caused by small targets and clutter interference. BRIEF DESCRIPTION OF DRAWINGS
[0034] Figure 1 The flowchart of the polarization SAR ship target detection method based on the multi-channel network;
[0035] Figure 2 Logical structure diagram of each module of the polarimetric SAR ship target detection system;
[0036] Figure 3 Part of the image of the data set used in the application;
[0037] Figure 4 Detection result map of the multi-channel network of the application;
[0038] Figure 5 Typical PR curve of the original YOLOv8 network for polarimetric SAR ship targets;
[0039] Figure 6 Typical PR curve of the multi-channel network of the application for polarimetric SAR ship targets. DETAILED DESCRIPTION
[0040] The technical solutions of the application are specifically described below in conjunction with the accompanying drawings of the specification. It should be noted that, in this document, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device.
[0041] The application provides a polarimetric SAR ship target detection method based on a multi-channel network, specifically as shown in Figure 1 , including the following steps:
[0042] S1. Construct a polarimetric SAR ship target detection system based on a multi-channel network; the polarimetric SAR ship target detection system based on a multi-channel network includes a data preprocessing module, a feature fusion module, a backbone feature extraction module, a pyramid feature extraction module, and a detection output module. By changing the YOLOv8 network structure to a dual-channel structure, the improved network can deeply mine and fully utilize the polarization characteristics of polarimetric SAR ship images, and better improve the polarimetric SAR ship target detection accuracy.
[0043] In specific embodiments, the logic structure diagram of each module of the polarimetric SAR ship target detection system is specifically shown in FIG. 2. The data preprocessing module is responsible for labeling the original data set, dividing it into a training set and a test set, and randomly rearranging it to improve the adaptability and generalization performance of the network. The feature fusion module focuses on integrating different feature information. The main channel and the auxiliary channel respectively pass through independent convolution blocks for shallow feature extraction, and the output feature maps are spliced in the channel dimension to generate fused high-order features; the shallow network (such as the first few layers of convolution) of the backbone feature extraction module focuses on local details, extracts edge, texture and other low-level features, and the deep network (such as the residual module) captures the global outline, spatial layout and class semantic information of the target by expanding the receptive field, and distinguishes ships from background clutter; the pyramid feature extraction module adopts the FPN (Feature Pyramid Network) structure, which fuses the semantic information of the high layer and the positioning accuracy of the low layer from top to bottom (high layer→low layer) of the backbone network output; the detection output module outputs the class probability of each anchor box through the convolution layer, predicts the offset of the bounding box and evaluates the probability of the existence of the target in the predicted box.
[0044] S2. Constructing a data set for training a polarimetric SAR ship target detection system based on a multi-channel network, wherein the data set set in this embodiment is the public FPSD polarimetric SAR ship target data set, and part of the image of the data set is specifically shown in FIG. Figure 3 Figure 3 wherein (a) is the image of data set sample 1, (b) is the image of data set sample 2, (c) is the image of data set sample 3, and (d) is the image of data set sample 4, and then polarimetric data extraction, data labeling and data enhancement are performed.
[0045] Wherein, S2 includes the following specific steps:
[0046] S21. Data preprocessing: extracting the main diagonal elements T11, T22, T33 of the coherence matrix of the polarimetric SAR data as the main channel data input, and performing eight-component decomposition on the coherence matrix to obtain the auxiliary channel data input;
[0047] In specific embodiments, specifically: the main diagonal elements T11, T22, T33 of the coherence matrix input as polarimetric SAR data are extracted as main channel input data, and the three channels correspond to the scattering energy of different polarization combinations (such as HH, HV, and VV), which can reflect the strong scattering characteristics of the target; the coherence matrix is decomposed into eight components to obtain auxiliary features including surface scattering, volume scattering, even scattering, helical scattering, rotational dihedral scattering, rotational dipole scattering, rotational 1 / 4 wave device scattering, and mixed dipole scattering, which constitute auxiliary channel input data. Such features can enhance the characterization ability of complex scattering mechanisms (such as ship and sea clutter mixed scenes).
[0048] S22. Constructing a data set: labeling target bounding boxes and class labels, and dividing training set and test set. In specific embodiments, the main channel input data and the auxiliary channel input data are uniformly labeled (bounding box and class label), and divided into training set and test set in the ratio of 7:3 to ensure the objectivity of model evaluation, avoid local overfitting in the training process by randomly shuffling sample order, and thus enhance the generalization and robustness of the model.
[0049] (4) Avoiding local overfitting in the training process by randomly shuffling sample order, thereby enhancing the generalization and robustness of the model.
[0050] S3. Training the data preprocessing module, feature fusion module, backbone feature extraction module, pyramid feature extraction module, detection output module, and class discrimination and positioning module of the polarimetric SAR ship target detection system based on the multi-channel network using the training set, and saving the weight parameters of the trained feature fusion module, backbone feature extraction module, pyramid feature extraction module, and detection output module;
[0051] S3 includes the following steps:
[0052] S31. Initialize the weight parameters: initialize the network weight parameters of the feature fusion module, backbone feature extraction module, pyramid feature extraction module, and detection output module using a normal distribution with a mean of 0 and a variance of 0.01;
[0053] S32. Set the training parameters of the polarimetric SAR ship target detection system based on the multi-channel network: use AdamW as the model training optimizer, set the hyperparameters β1 of the model training optimizer to 0.95, β2 to 0.9, and the weight decay to 5×10 -4 The batch size is 16, and the maximum training step is 200;
[0054] S33. Calculate the loss function value Loss, which includes the coordinate prediction loss function, the confidence loss function, and the class prediction loss function;
[0055] S34. The termination condition for determining whether the training is finished is that the loss function value Loss < 0.005 or 200 training rounds are reached.
[0056] In a specific embodiment, the process is as follows:
[0057] (1) Initialize the weight parameters. The network weight parameters of the feature fusion module, the backbone feature extraction module, the pyramid feature extraction module, and the detection output module are initialized using a normal distribution with a mean of 0 and a variance of 0.01.
[0058] (2) Set the training parameters of the polarimetric SAR ship target detection system based on the multi-channel network. Set the batch size (batch_size) to 16 and the maximum training step (maxepoch) to 200. Use AdamW as the model training optimizer, and set the hyperparameters β1 of the model training optimizer to 0.95, β2 to 0.9, and the weight decay to 0.001.
[0059] (3) Initialize the training iteration parameter epoch = 1, the batch iteration parameter num_batch = 0 (indicating that the current batch is the num_batchth batch in the current epoch), and the error value of the first round of training Loss = 0.
[0060] (4) The feature fusion module extracts the preprocessed polarimetric SAR ship target data, fuses the features of the two channels, and then inputs them to the backbone feature extraction module.
[0061] (5) The backbone feature extraction module extracts the edge, texture, and other low-level features of the polarimetric SAR ship target data, as well as the global contour, spatial layout, and class semantic information, and then inputs them to the pyramid feature extraction module.
[0062] (6) The pyramid feature extraction module uses the FPN structure to fuse the high-level semantic information and the low-level positioning accuracy of the multi-scale features output by the backbone network from top to bottom.
[0063] (7) The detection output module calculates the loss function value Loss through the coordinate prediction loss function, the confidence prediction loss function, and the class prediction loss function.
[0064] (8) Update the network weight parameters of the polarimetric SAR ship target detection system based on the multi-channel network using gradient backpropagation.
[0065] (9) When num_batch is equal to 29, epoch is incremented by 1.
[0066] (10) determining whether the training is completed, if the Loss calculated in (7) is less than or equal to 0.005, the training is completed; if the epoch is less than maxepoch and the Loss calculated in (7) is greater than 0.005, the training is continued in (4); if the epoch is equal to maxepoch, it is indicated that the training is completed, and the network weight parameters of the feature fusion module, the backbone feature extraction module, the pyramid feature extraction module and the detection output module after the current epoch are saved.
[0067] S4. The polarimetric SAR ship target detection system based on the multi-channel network loads the pre-stored weight parameters to obtain a final detection system.
[0068] S5. The final detection system detects the input polarimetric SAR ship target by using the feature fusion module, the backbone feature extraction module, the pyramid feature extraction module and the detection output module, and outputs the result.
[0069] The application further provides a polarimetric SAR ship target detection device based on a multi-channel network, which comprises a memory for storing a pre-processing data set and network weight parameters, a processor configured to perform data preprocessing, dual-channel feature fusion of a main channel and an auxiliary channel, multi-scale feature extraction and target detection output, and an output interface for transmitting a detection result of a target category, a bounding box coordinate and a confidence.
[0070] After strict testing, the trained network of the application achieves an average detection accuracy of 92.3% on the test set, and exhibits excellent convergence stability. This result verifies the excellent performance of the application in polarimetric SAR ship target detection.
[0071] The application is verified by simulation tests to have good polarimetric SAR ship target detection effect, and typical detection results are as shown in Figure 4 , Figure 4 wherein (a) is an image of a main channel detection result 1, (b) is an image of an auxiliary channel detection result 1, (c) is an image of a main channel detection result 2, and (d) is an image of an auxiliary channel detection result 2. Figure 4 It can be seen that the polarimetric SAR ship target detection method based on the multi-channel network of the application not only accurately detects ships in a complex scene, but also accurately detects ship targets in a multi-ship target scene. In addition, the PR (precision-recall) curve of the original YOLOv8 network is as shown in Figure 5 , and the PR curve of the improved YOLOv8 network in the application is as shown in Figure 6 . Figure 5 , Figure 6It can be seen that, compared with the original YOLOv8 network, the detection effect of the polarimetric SAR ship target detection method based on the multi-channel network of the application is improved by 1.1 times, which fully verifies the practicability of the polarimetric SAR ship target detection method based on the multi-channel network of the application.
[0072] Finally, it should be pointed out that, although the present application has been described with reference to the current specific embodiments, those skilled in the art should realize that the above embodiments are only used to illustrate the present application, and are not used as a limitation on the present application, and various equivalent changes or replacements can be made without departing from the concept of the present application, therefore, any changes or modifications of the above embodiments within the scope of the spirit of the present application will fall within the scope of the claims of the present application.
Claims
1. Polarimetric SAR ship target detection method based on multi-channel network, characterized by: The following steps are involved: S1. Build a polarimetric SAR ship target detection system based on a multi-channel network; S2. Construct a dataset for training a multi-channel network-based polarimetric SAR ship target detection system. S3. Use the training set to optimize the weights of the multi-channel network-based polarimetric SAR ship target detection system and save the trained weight parameters; S4. Loading the pre-stored weight parameters into the multi-channel network-based polarimetric SAR ship target detection system to obtain the final detection system; S5. The final detection system detects the input polarization SAR ship target and outputs the result.
2. The polarization SAR ship target detection method based on a multi-channel network according to claim 1, characterized in that: The polarimetric SAR ship target detection system in S1 includes: The preprocessing module labels the original dataset, splits it into training and test sets, and randomly rearranges them to improve the adaptability and generalization performance of the network; and The feature fusion module integrates different feature information. The main channel and auxiliary channel are respectively extracted through independent convolution blocks. The output feature maps are spliced in the channel dimension to generate fused high-order features; and The backbone feature extraction module uses shallow convolution to extract edge and texture features, and the deep residual module captures the global semantic information of the target; and The pyramid feature extraction module uses the FPN structure to fuse the multi-scale features of the backbone network from top to bottom, combining high-level semantic information with low-level positioning accuracy; and The detection output module outputs the category probability of each anchor box through the convolutional layer, predicts the offset of the bounding box, and evaluates the probability of the existence of the target in the predicted box.
3. The method for detecting ship targets using polarimetric SAR based on a multi-channel network according to claim 2, wherein: The S2 comprises the following steps: S21. Data preprocessing: Extract the main diagonal elements T11, T22, and T33 of the coherence matrix of the polarimetric SAR data as the main channel data input. Perform an eight-component decomposition of the coherence matrix to obtain the auxiliary channel data input. S22. Build a dataset: label the target bounding boxes and category labels, and divide the dataset into training and test sets.
4. The method for detecting ship targets using polarimetric SAR based on a multi-channel network according to claim 3, wherein: The main channel data input in S21 is the main diagonal elements T11, T22, and T33 components of the coherence matrix; the auxiliary channel data input is the surface scattering, volume scattering, even scattering, spiral scattering, rotating dihedral angle scattering, chiral dipole scattering, chiral 1 / 4 wave device scattering and mixed dipole scattering characteristics obtained by eight-component decomposition.
5. The method for detecting ship targets using polarimetric SAR based on a multi-channel network according to claim 4, wherein: The S3 includes the following steps: S31. Initialize weight parameters: Use a normal distribution with a mean of 0 and a variance of 0.01 to initialize the network weight parameters of the feature fusion module, backbone feature extraction module, pyramid feature extraction module, and detection output module; S32. Set the training parameters of the polarimetric SAR ship target detection system based on a multi-channel network: AdamW is used as the model training optimizer, and the hyperparameters β1 and β2 of the model training optimizer are set to 0.95 and 0.9, and the weight decay is 5×10 -4 The batch size is 16 and the maximum training step is 200; S33. Calculate the loss function value Loss, which includes the coordinate prediction loss function, the confidence loss function, and the category prediction loss function; S34. The termination condition for determining whether the training is completed is that the loss function value Loss ≤ 0.005 or 200 training rounds are reached.
6. Polarimetric SAR ship target detection device based on multi-channel network, characterized in that: include: Memory: stores preprocessed data sets and network weight parameters; and Processor: configured to perform data preprocessing, dual-channel feature fusion of the main channel and the auxiliary channel, multi-scale feature extraction and target detection output; and Output interface: transmits the detection results of target category, bounding box coordinates and confidence.
7. The polarization SAR ship target detection device based on a multi-channel network according to claim 6, characterized in that: Its processor architecture adopts an improved YOLOv8 network structure, which expands a single input channel into a dual-channel structure.