A ship detection method based on polarization characteristic driven full polarization SAR image

Abstract

The application discloses a full-polarization SAR image ship detection method based on polarization characteristic driving, comprising the following steps: S1, preprocessing a full-polarization SAR image to obtain polarization characteristic data of the full-polarization SAR image; S2, constructing a polarization characteristic driven convolutional neural network, inputting the polarization characteristic data into a backbone network and a dense structure to extract polarization characteristics contained in the polarization characteristic data, and performing multi-scale fusion in a feature pyramid, wherein the backbone network and the dense structure are used for context feature reuse in the network; S3, training the polarization characteristic driven convolutional neural network according to preset training parameters and a loss function to form a full-polarization SAR image target detection network, and using the trained polarization characteristic driven full-polarization SAR image target detection network to perform full-polarization SAR image ship detection and index evaluation. The application can effectively improve the detection capability of small targets in a scene and effectively detect ships in a strong scattering background.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and synthetic aperture radar target detection technology, and more specifically, to a method for ship detection based on polarization characteristics-driven fully polarimetric SAR images. Background Technology

[0002] With the development of satellite technology, Synthetic Aperture Radar (SAR) plays a vital role in Earth observation due to its unique all-weather and all-day advantages, such as land cover classification, change detection, disaster loss assessment, maritime traffic control, and surface parameter inversion. Various countries have launched SAR satellites, including China's Gaofen-3, Japan's ALOS, Europe's Sentinel-1, and Canada's RADARSAT-2. SAR holds a crucial position in Earth observation research, and ship detection at sea is a very important application of this technology.

[0003] For decades, threshold-based detection methods have been the mainstream approach for ship detection in the SAR field, such as the classic constant false alarm rate (CFAR) algorithm. The key to threshold-based detection methods lies in the selection of the distribution model and the estimation of its parameters. Commonly used distribution models include Gaussian, Weibull, K-distribution, generalized K-distribution, and generalized gamma distribution. Among these, the generalized gamma distribution achieves higher fitting accuracy compared to other distribution models and is more suitable for clutter fitting in SAR images.

[0004] In recent years, with the continuous development of satellite technology, from the initial single-polarization mode to the current multi-polarization mode, SAR images contain increasingly rich information, especially in fully polarimetric SAR data, which contains abundant polarization information and has played a significant advantage in ship detection. Convolutional neural networks (CNNs) have been widely studied in the field of computer vision due to their powerful feature extraction and expressive capabilities. In the field of target detection, neural networks such as SSD, Faster-RCNN, YOLO, and RetainNet have achieved remarkable results. Deep learning has also achieved great success in the field of computer vision and has attracted the attention of researchers in the SAR field. Utilizing existing public datasets, many scholars have proposed corresponding CNNs and achieved significant detection results. However, the application of convolutional neural networks to ship target detection in (Pol)SAR still faces the following problems: (1) Due to the unique imaging mechanism of synthetic aperture radar, targets in SAR images appear as bright areas, which are easily affected by strong scattering backgrounds, making it difficult for existing algorithms to extract useful features; (2) Existing SAR image neural network ship detection algorithms use single-channel intensity data as input, ignoring the polarization characteristics in SAR images; (3) SAR images contain targets of various scales and many small targets, posing a significant challenge to the detection capabilities of neural network algorithms. In summary, relying solely on the intensity data of SAR images to construct features limits the performance of convolutional neural network algorithms. Therefore, there is an urgent need for a convolutional neural network that can be driven by polarization characteristic data to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a ship detection method based on polarization characteristics driven by fully polarimetric SAR images, so as to overcome the defects of the existing technology.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A ship detection method based on polarization characteristics driven by fully polarimetric SAR images includes the following steps:

[0008] S1. Preprocess the fully polarimetric SAR image to obtain the polarization characteristic data of the fully polarimetric SAR image;

[0009] S2. Construct a polarization characteristic-driven convolutional neural network, input polarization characteristic data into the backbone network and dense structure to extract the polarization characteristics contained in the polarization characteristic data, and perform multi-scale fusion in the feature pyramid. The backbone network and dense structure are used for context feature reuse in the network.

[0010] S3. According to the preset training parameters and loss function, the polarization characteristic-driven convolutional neural network is trained to form a fully polarimetric SAR image target detection network, and the trained polarization characteristic-driven fully polarimetric SAR image target detection network is used to perform fully polarimetric SAR image ship detection and index evaluation.

[0011] Furthermore, step S1 specifically includes the following steps:

[0012] S10. Calculate the wave polarization anisotropy ΔS in a fully polarimetric SAR image using the polarization scattering matrix S. n ;

[0013] S11. Using the covariance matrix of the fully polarimetric SAR image, perform four-component decomposition to obtain the specular scattering P. s Double-hop scattering P d Volume scattering P v and helical scattering P h The power value;

[0014] S12. Obtain the polarization characteristic data of the input network using wave polarization anisotropy and the power value of the scattering model.

[0015] Furthermore, the specific calculation formula for obtaining the polarization characteristic data in step S12 is as follows:

[0016]

[0017] Furthermore, the construction of the polarization-driven convolutional neural network in step S2 specifically includes the following steps:

[0018] S20. Input the polarization characteristic data into the backbone network and use the backbone network to extract the features contained in the polarization characteristics.

[0019] S21. Input the features into a dense structure and reuse the polarization features;

[0020] S22. Input the features processed by the dense structure into the feature pyramid and perform feature map polarization characteristic fusion at different scales.

[0021] Furthermore, the polarization characteristic data of the input backbone network in step S20 is in the following form:

[0022]

[0023] Furthermore, in the dense structure of step S21, a given single sample x0 is processed by network convolution. The network contains L layers, and each layer is a nonlinear transformation H(.). Let the output of the i-th layer be x. i Then the feature map received by the i-th layer in the dense structure is:

[0024] X i =H l ([X0, X1, ..., X i-1 ])

[0025] In the formula, [X0, X1, ..., X i-1 ] represents the connection of feature maps generated from layers 0, ..., i-1, x i It incorporates X0…X i-1 Features within the layer.

[0026] Compared with the prior art, the advantages of the present invention are as follows: The present invention provides a ship detection method based on polarization characteristics driven by fully polarimetric SAR images, which makes full use of the polarization characteristics of fully polarimetric SAR images, effectively suppresses interference such as ghosting and side lobes in the scene, and greatly improves the network's ability to detect targets; on this basis, by using feature pyramids to fuse polarization characteristic data at different scales, small targets contained in the scene can be effectively extracted. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is the overall architecture diagram of the present invention.

[0029] Figure 2 This invention is a fully polarimetric SAR image target detection network driven by polarization characteristics.

[0030] Figure 3 This is the receptive field module in this invention.

[0031] Figure 4 This is the detection result of the LS-SSDD-v1 dataset in this invention.

[0032] Figure 5 These are the polarization characteristic data SA and the detection results of the common polarization channel VV in this invention. Detailed Implementation

[0033] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0034] See Figure 1 and Figure 2As shown, this embodiment discloses a ship detection method based on polarization characteristics driven by fully polarimetric SAR images, including the following steps:

[0035] Step S1: Perform data preprocessing on the fully polarimetric SAR image to obtain the polarization characteristics in the fully polarimetric SAR image.

[0036] Specifically, step S1 includes the following steps:

[0037] Step S10: Using C-band fully polarimetric SAR image data acquired by the Chinese Gaofen-3 satellite, a two-dimensional complex scattering matrix [S] is used to characterize the information of each pixel unit in the image. The formula for the polarimetric scattering matrix of a single pixel is:

[0038]

[0039] The Muller matrix is ​​derived from this scattering matrix, and finally the wave polarization anisotropy ΔS of the fully polarimetric SAR data is obtained. n .

[0040] Step S11: Perform four-component decomposition using the covariance matrix of the fully polarimetric SAR data to obtain the specular scattering P. s Double-hop scattering P d Volume scattering P v and helical scattering P h The power is expressed as follows:

[0041]

[0042] Step S12, export the polarization characteristic data of a single pixel, the expression of which is as follows:

[0043]

[0044] Step S2: Construct a fully polarimetric SAR image target detection network driven by polarization characteristics, which mainly includes three modules: a dense module, a receptive field module, and a feature fusion module.

[0045] Step S2 specifically includes the following steps:

[0046] S20. Input the polarization characteristic data into the backbone network and use the backbone network to extract the features contained in the polarization characteristics.

[0047] The polarization characteristic data input to the neural network is in the following form:

[0048]

[0049] S21. Input the features into the dense structure and reuse the polarized features to ensure that the feature information is not lost during the transmission of the network structure.

[0050] S22. Input the features processed by the dense structure into the feature pyramid and perform feature map polarization characteristic fusion at different scales.

[0051] like Figure 2 As shown, after feature extraction from the backbone network, the polarization characteristic data is input into a dense structure. This structure utilizes dense connections to enhance feature propagation, achieve feature reuse, and significantly reduce the number of network parameters. Assume a single sample x0 is processed by the network's convolutional layer, where the network contains L layers, each a nonlinear transformation H(.). Let the output of the i-th layer be x. i Then the feature map received by the i-th layer in the dense structure is:

[0052] X i =H l ([X0, X1, ..., X i-1 ])

[0053] In the formula, [X0, X1, ..., X i-1 ] represents the connection of feature maps generated from layers 0, ..., i-1, x i It incorporates X0…X i-1 Features within the layer.

[0054] like Figure 3 As shown, the receptive field module incorporates dilated convolutions, effectively increasing the network's receptive field and enabling the acquisition of richer features. The receptive field module has two main characteristics: 1) convolutional layers with kernels of different sizes form a multi-branch structure; 2) the introduction of dilated convolutional layers effectively increases the receptive field.

[0055] like Figure 2 As shown, the features extracted through the dense structure are input into the feature pyramid for fusion. Since targets in SAR images have multi-scale characteristics, a feature pyramid is used to fuse multi-scale information about targets in SAR images. In neural networks, shallow features contain more low-level, simple features, such as edge features; while deep neural networks contain more semantic features. Images contain targets of different sizes, and different targets have different features. Shallow features can distinguish simple targets; deep features can distinguish complex targets. Fusing shallow features with semantic features enables multi-scale target detection.

[0056] Step S3: Train the polarization-driven convolutional neural network according to the preset training parameters and loss function to form a fully polarimetric SAR image target detection network, and use the trained polarization-driven fully polarimetric SAR image target detection network to perform fully polarimetric SAR image ship detection and index evaluation.

[0057] In this embodiment, to illustrate the superiority of this network, tests are first conducted using LS-SSDD-v1 data, such as... Figure 4 As shown. For small target detection, the network in this embodiment outperforms the three networks compared. Furthermore, to illustrate the effectiveness of using polarization characteristic data as network input, the polarization characteristic data and the power value of the common polarization channel VV are compared, and the comparison results are shown below. Figure 5 As shown. In addition to testing the network in this embodiment, the network compared in this paper was also used for testing. The results show that using polarization characteristic data as network input yields better detection results than using the common polarization channel VV as network input.

[0058] This invention utilizes polarimetric characteristic data from fully polarimetric SAR images as network input. As mentioned earlier, relying solely on target intensity information for ship detection cannot effectively suppress clutter backgrounds and cannot provide a basis for inference between background clutter and ships from the data source. Therefore, this paper uses polarimetric characteristic data as input data for a neural network. Preprocessed polarimetric characteristic data can effectively suppress clutter and highlight targets, providing rich target features for the neural network. Furthermore, considering the issue of small target scales in SAR images, a feature pyramid with a receptive field module is proposed, which can better capture the details of small targets and improve the detection accuracy of small targets. Through the above two analyses, this invention uses polarimetric characteristic data as network input to effectively suppress clutter backgrounds and highlight targets in complex scenes, greatly improving the network's judgment ability; based on this, a feature pyramid with a receptive field module is used to capture the features of small targets, thereby improving the detection capability of small targets.

[0059] Although embodiments of the present invention have been described in conjunction with the accompanying drawings, the patent owner may make various modifications or alterations within the scope of the appended claims, as long as they do not exceed the protection scope described in the claims of the present invention, they shall be within the protection scope of the present invention.

Claims

1. A ship detection method based on polarization characteristics driven by fully polarimetric SAR images, characterized in that, Includes the following steps: S1. Preprocess the fully polarimetric SAR image to obtain the polarization characteristic data of the fully polarimetric SAR image; S2. Construct a polarization characteristic-driven convolutional neural network, input polarization characteristic data into the backbone network and dense structure to extract the polarization characteristics contained in the polarization characteristic data, and perform multi-scale fusion in the feature pyramid. The backbone network and dense structure are used for context feature reuse in the network. S3. According to the preset training parameters and loss function, the polarization characteristic-driven convolutional neural network is trained to form a fully polarimetric SAR image target detection network, and the trained polarization characteristic-driven fully polarimetric SAR image target detection network is used to perform fully polarimetric SAR image ship detection and index evaluation. Step S1 specifically includes the following steps: S10. Obtain the wave polarization anisotropy in a fully polarimetric SAR image using the polarization scattering matrix S. ; S11. Using the covariance matrix of the fully polarimetric SAR image, perform four-component decomposition to obtain the specular scattering... Double-hop scattering Volume scattering and spiral scattering The power value; S12. Obtain the polarization characteristic data of the input network using wave polarization anisotropy and the power value of the scattering model.

2. The ship detection method based on polarization characteristics driven by fully polarimetric SAR images according to claim 1, characterized in that, The specific calculation formula for obtaining polarization characteristic data in step S12 is as follows: 。 3. The ship detection method based on polarization characteristics driven by fully polarimetric SAR images according to claim 1, characterized in that, The construction of the polarization-driven convolutional neural network in step S2 specifically includes the following steps: S20. Input the polarization characteristic data into the backbone network and use the backbone network to extract the features contained in the polarization characteristics. S21. Input the features into a dense structure and reuse the polarization features; S22. Input the features processed by the dense structure into the feature pyramid and perform feature map polarization characteristic fusion at different scales.

4. The ship detection method based on polarization characteristics driven by fully polarimetric SAR images according to claim 3, characterized in that, The polarization characteristic data of the backbone network input in step S20 is in the following form: 。 5. The ship detection method based on polarization characteristics driven by fully polarimetric SAR images according to claim 1, characterized in that, In the dense structure of step S21, a given single sample is set. After convolutional processing, the network contains L layers, each of which is a nonlinear transformation H(.). Let the output of the i-th layer be... Then the feature map received by the i-th layer in the dense structure is: In the formula, Indicates from The connection of feature maps generated in the layer, Fusion Features within the layer.

Images