SAR image scattering feature extraction method based on self-supervised learning

Abstract

The invention discloses an SAR image scattering feature extraction method based on self-supervised learning. The method comprises the following steps: inputting an original SAR image into a scattering feature extraction network; performing feature extraction on the original SAR image by using a convolution module to obtain a first feature map; performing maximum pooling processing on the first feature map, and inputting the first feature map into a residual module to obtain a second feature map; the residual module comprises a plurality of residual units, and each residual unit is used for generating a first sub-feature map according to the output of the last-level residual unit; inputting the second feature map into a convolution attention mechanism module, and generating a third feature map containing channel dimension feature information and spatial dimension feature information; inputting the third feature map into a processing module, and combining the first feature map and the plurality of first sub-feature maps to generate a fourth feature map; and respectively inputting the fourth feature map into three detection heads, and performing parameter prediction by the three detection heads based on activation functions corresponding to the three detection heads to obtain scattering features of the original SAR image.

Description

Technical Field

[0001] The invention belongs to the technical field of SAR image processing, and in particular relates to a SAR image scattering feature extraction method based on self-supervised learning. Background Art

[0002] Synthetic Aperture Radar (SAR) has the characteristics of all-day and all-weather, and can achieve high-precision observation of the ground without being affected by the time and space environment. SAR images are widely used in civil and military fields, such as natural resource monitoring, target detection and target recognition. However, the existing SAR image applications are mainly based on the amplitude value of SAR images, ignoring the physical information implicit in SAR images, such as scattering characteristics. In the scattering center theory, when the object size is greater than ten times the radar emission wavelength, the radar works in the optical region. The electromagnetic scattering of the target originates from the combination of various parts of the target. These local scattering sources are called scattering centers, which convey the structural information of the target. The scattering feature extraction problem of SAR images is essentially the parameter estimation problem of the scattering center model. By selecting a suitable scattering center model, the SAR image can be regarded as a combination of scattering center models. The goal of scattering feature extraction is to estimate the parameters of the scattering center model so that the image reconstructed based on these parameters remains the same as the original SAR image. The original SAR image can be decomposed into a combination of multiple scattering center models. The estimated parameters can be used for tasks such as SAR target recognition.

[0003] In the prior art, the methods for extracting scattering features of SAR images mainly include: a method based on estimating two-dimensional poles and amplitude coefficients in the Prony model, a radar echo feature extraction method based on the attribute scattering center model, an attribute scattering center feature extraction method based on regional decoupling, and an estimation method of scattering center parameter decoupling based on the improved RELAX algorithm. Although the above methods have achieved the task of scattering feature extraction, most of these methods are iterative estimation methods based on frequency domain or image domain, which need to face complex initial value setting problems, and require large amount of calculation and long calculation time, which limits the application of scattering features.

[0004] In addition, there is also a SAR image scattering center extraction method based on a supervised learning neural network in the prior art. This method involves a traditional supervised learning process and requires providing labels of scattering features, but it is very difficult to obtain these labels. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a SAR image scattering feature extraction method based on self-supervised learning. The technical problem to be solved by the present invention is achieved through the following technical solutions: The present invention provides a SAR image scattering feature extraction method based on self-supervised learning, comprising: Get the original SAR image; The original SAR image is input into a scattering feature extraction network, so that the scattering feature extraction network performs the following steps: Using a convolution module to extract features from the original SAR image to obtain a first feature map; The first feature map is subjected to maximum pooling processing and then input into a residual module to obtain a second feature map; the residual module includes a plurality of residual units connected in sequence, each of which is used to generate a first sub-feature map according to an output of a residual unit at a previous level; Inputting the second feature map into a convolutional attention mechanism module to generate a third feature map including channel dimension feature information and spatial dimension feature information; The third feature map is input into a processing module, so that a plurality of sequentially connected processing units in the processing module generate a third sub-feature map according to two-way input data of the processing unit; wherein the two-way input data of the first-level processing unit in the processing module include: the third feature map and the first sub-feature map generated by the residual unit corresponding to the processing unit, the two-way input data of the last-level processing unit include: the third sub-feature map generated by the previous-level processing unit and the first feature map, and the two-way input data of the remaining processing units include: the third sub-feature map generated by the previous-level processing unit and the first sub-feature map generated by the residual unit corresponding to the processing unit; The third sub-feature map generated by the last-level processing unit is used as the fourth feature map, and is respectively input into the first detection head, the second detection head and the third detection head, and the first detection head, the second detection head and the third detection head perform parameter prediction based on their corresponding activation functions to obtain the scattering characteristics of the original SAR image.

[0006] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention introduces a convolutional attention mechanism module into the scattering feature extraction network, which can fully extract the channel information and spatial information of the original SAR image, help the scattering feature extraction network to identify and enhance the scattering feature information of the input original SAR image, and capture long-range dependencies.

[0007] (2) The present invention uses different detection heads to predict different scattering characteristic parameters, thereby avoiding mutual interference between different scattering characteristic parameters in the same attribute parameter graph.

[0008] (3) In each round of training of the scattering feature extraction network, a reconstructed SAR image is generated using a pre-constructed simplified attribute scattering center model based on the predicted scattering feature parameters, namely the echo amplitude, length and azimuth of each scattering center in the original SAR image. The reconstructed SAR image is then incorporated into the calculation of the loss function value. The network parameters of the scattering feature extraction network are optimized based on the idea of ​​minimizing the difference between the reconstructed SAR image and the original SAR image, thereby achieving self-supervised learning without the need for pre-labeled labels. In addition, when constructing the simplified attribute scattering center model, the complex original attribute scattering center model is simplified, thereby improving the image reconstruction speed while ensuring the image reconstruction accuracy.

[0009] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 is a flow chart of a SAR image scattering feature extraction method based on self-supervised learning provided by an embodiment of the present invention; Figure 2 is a schematic diagram of a scattering feature extraction network provided by an embodiment of the present invention; Figure 3 is a schematic diagram of a convolutional attention mechanism module provided by an embodiment of the present invention; Figure 4 is a schematic diagram of a channel attention unit provided by an embodiment of the present invention; Figure 5 is a schematic diagram of a spatial attention unit provided by an embodiment of the present invention; Figure 6 is a schematic diagram of a first detection head, a second detection head and a third detection head provided by an embodiment of the present invention; Figure 7 is a schematic diagram of the geometric relationship between a target and a radar provided by an embodiment of the present invention; Figure 8 It is the SAR image of 10 types of ground vehicle targets in the MSTAR dataset; Fig. 9 It is the loss value curve of the scattering feature extraction network on the training set and the data set; Fig.10 This is a comparison chart of the reconstruction results of the original SAR image containing the type 2S1 ground vehicle target; Fig.11 This is a comparison chart of the reconstruction results of the original SAR image containing the type D7 ground vehicle target; Fig.12 This is a comparison chart of the reconstruction results of the original SAR image containing the T62 ground vehicle target; Fig.13This is a comparison chart of the reconstruction results of the original SAR image containing the ZIL131 ground vehicle target; Fig.14 This is a comparison chart of the reconstruction results of the original SAR image containing the ZSU23 / 4 ground vehicle target. DETAILED DESCRIPTION

[0011] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0012] Figure 1 is a flow chart of a SAR image scattering feature extraction method based on self-supervised learning provided by an embodiment of the present invention, Figure 2 FIG. 1 is a schematic diagram of a scattering feature extraction network provided by an embodiment of the present invention. Figure 1~2 As shown, an embodiment of the present invention provides a SAR image scattering feature extraction method based on self-supervised learning, comprising: S1. Obtain the original SAR image.

[0013] In this embodiment, the real part, imaginary part and amplitude value of the original SAR image can be spliced ​​and input into the scattering feature extraction network.

[0014] S2. Input the original SAR image into the scattering feature extraction network, so that the scattering feature extraction network performs the following steps: S201, using a convolution module to perform feature extraction on the original SAR image to obtain a first feature map; S202, performing maximum pooling processing on the first feature map and inputting the resultant into a residual module to obtain a second feature map; the residual module includes a plurality of residual units connected in sequence, each residual unit being used to generate a first sub-feature map according to an output of a residual unit at a previous level; S203, inputting the second feature map into a convolutional attention mechanism module to generate a third feature map including channel dimension feature information and spatial dimension feature information; S204, inputting the third feature map into the processing module, so that a plurality of sequentially connected processing units in the processing module generate a third sub-feature map according to their own two-way input data; wherein the two-way input data of the first-level processing unit in the processing module include: the third feature map and the first sub-feature map generated by the residual unit corresponding to itself, the two-way input data of the last-level processing unit include: the third sub-feature map generated by the previous-level processing unit and the first feature map, and the two-way input data of the remaining processing units include: the third sub-feature map generated by the previous-level processing unit and the first sub-feature map generated by the residual unit corresponding to itself; S205. Use the third sub-feature map generated by the last-level processing unit as the fourth feature map, and input the third sub-feature map into the first detection head, the second detection head and the third detection head respectively. The first detection head, the second detection head and the third detection head perform parameter prediction based on their corresponding activation functions to obtain the scattering characteristics of the original SAR image.

[0015] Please continue to see Figure 2 The residual module includes a first residual unit, a second residual unit, a third residual unit and a fourth residual unit, and the processing module includes a first processing unit, a second processing unit, a third processing unit and a fourth processing unit. It should be noted that: Figure 2 The four processing units included in the processing module are named as the first processing unit, the second processing unit, the third processing unit and the fourth processing unit from top to bottom. However, considering the actual data processing flow, this embodiment uses the fourth processing unit as the first-level processing unit in the processing module, and the first processing unit is the last-level processing unit in the processing module.

[0016] After the step of extracting features from the original SAR image using a convolution module to obtain a first feature map, the method further includes: inputting the first feature map into a first processing unit; Optionally, in step S202, the step of performing maximum pooling processing on the first feature map and inputting the processed feature map into a residual module to obtain a second feature map includes: The first residual unit generates the first sub-feature map based on the first feature map after the maximum pooling process. , and the first sub-feature map input into the second processing unit and the second residual unit respectively; The second residual unit is based on the first sub-feature map Generate the first sub-feature map , and the first sub-feature map input into the third processing unit and the third residual unit respectively; The third residual unit is based on the first sub-feature map Generate the first sub-feature map , and the first sub-feature map input into the fourth processing unit and the fourth residual unit respectively; The fourth residual unit is based on the first sub-feature map Generate the first sub-feature map , and the first sub-feature map As the second feature map, it is input into the convolutional attention mechanism module.

[0017] Figure 3 is a schematic diagram of a convolutional attention mechanism module provided by an embodiment of the present invention, Figure 4 is a schematic diagram of a channel attention unit provided in an embodiment of the present invention, Figure 5 is a schematic diagram of a spatial attention unit provided by an embodiment of the present invention. Figure 3 As shown, the convolutional attention mechanism module includes a channel attention unit and a spatial attention unit; In step S203, the second feature map is input into the convolutional attention mechanism module to generate a third feature map containing channel dimension feature information and spatial dimension feature information, including: The second feature map is input into the convolutional attention mechanism module, and the channel attention vector is extracted using the channel attention unit. , and based on the channel attention vector Weight the second feature map to generate the second sub-feature map ; The second sub-feature map Input the spatial attention unit and extract the spatial attention vector , and based on the spatial attention vector For the second sub-feature map Weighted, generate the third feature map.

[0018] Specifically, if Figure 4 As shown, the second feature map is input into the convolutional attention mechanism module, and the channel attention unit is used to extract the channel attention vector The steps include: The second feature map is input into the convolutional attention mechanism module, and the maximum pooling process and the average pooling process are performed respectively; The second feature map after the maximum pooling process and the second feature map after the average pooling process are respectively input into a shared multi-layer perceptron, where the shared multi-layer perceptron is a two-layer fully connected network; The channel attention vector is obtained by adding the two outputs of the shared multilayer perceptron and processing them with the Sigmoid activation function. .

[0019] Furthermore, if Figure 5 As shown, the second sub-feature map Input the spatial attention unit and extract the spatial attention vector The steps include: The second sub-feature map Input the spatial attention unit, perform maximum pooling, average pooling and convolution in sequence, and then process it through the Sigmoid activation function to obtain the spatial attention vector .

[0020] Optionally, step S203 of inputting the third feature map into the processing module so that a plurality of sequentially connected processing units in the processing module generate a third sub-feature map according to their own two-path input data includes: The third feature map is input into the processing module, and the third feature map is upsampled by the fourth processing unit and combined with the first sub-feature map After splicing, convolution processing is performed to generate the third sub-feature map ; The third processing unit processes the third sub-feature map Upsample and compare with the first sub-feature map After splicing, convolution processing is performed to generate the third sub-feature map ; The second processing unit processes the third sub-feature map Upsample and compare with the first sub-feature map After splicing, convolution processing is performed to generate the third sub-feature map ; The first processing unit processes the third sub-feature map Up-sample and concatenate with the first feature map and perform convolution to generate the third sub-feature map As the fourth characteristic diagram.

[0021] Figure 6 Schematic diagram of the first detection head, the second detection head and the third detection head provided by the embodiment of the present invention. Figure 6 In step S205, the third sub-feature map generated by the last level processing unit is used as the fourth feature map, and is input into the first detection head, the second detection head and the third detection head respectively, and the first detection head, the second detection head and the third detection head perform parameter prediction based on their corresponding activation functions to obtain the scattering characteristics of the original SAR image, including: The third sub-feature map generated by the last-level processing unit is used as the fourth feature map, and the fourth feature map is input into the first detection head, the second detection head and the third detection head respectively; The first detection head generates a first attribute parameter map with the same scale as the original SAR image based on the ReLU activation function. The value of each point in the first attribute parameter map represents the predicted echo amplitude of each scattering center in the original SAR image. ; The second detection head generates a second attribute parameter map with the same scale as the original SAR image based on the improved Sigmoid activation function. The value of each point in the second attribute parameter map represents the length of each scattering center in the original SAR image. ; The third detection head generates a third attribute parameter map with the same scale as the original SAR image based on the improved Tanh activation function. The value of each point in the third attribute parameter map represents the predicted initial pointing angle of each scattering center in the original SAR image. .

[0022] In this embodiment, the first detection head, the second detection head and the third detection head respectively output the first attribute parameter map, the second attribute parameter map and the third attribute parameter map of the same scale as the original SAR image. Since the value range of the echo amplitude of the scattering center is greater than 0, the ReLU activation function is used to generate the first attribute parameter map, and the length of the scattering center is generally not more than 2 meters. Therefore, the improved Sigmoid activation function is used to generate the second attribute parameter map: , Represents the independent variable.

[0023] In addition, the azimuth angle range is , so the third detection head uses the improved Tanh activation function to generate the third attribute parameter map. Optionally, the expression of the improved Tanh activation function is: , Represents the independent variable.

[0024] It should be noted that in the above-mentioned SAR image scattering feature extraction method based on self-supervised learning, the scattering feature extraction network is trained based on the mean square error loss function; Among them, in each round of training of the scattering feature extraction network, the loss value of the mean square error loss function is: the mean of the sum of the squares of the differences between the pixel values ​​of the original SAR image and the reconstructed SAR image.

[0025] Optionally, generate a reconstructed SAR image according to the following steps: Based on the echo amplitude of each scattering center in the original SAR image predicted during this round of training, the scattering centers in the original SAR image are arranged in descending order, and the top N scattering centers; Before acquisition N The echo amplitude, length and azimuth of each scattering center are obtained and substituted into the pre-built simplified attribute scattering center model to generate a reconstructed SAR image.

[0026] In this embodiment, N =70, that is, the 70 scattering centers with the largest echo amplitudes are selected, and then the reconstructed SAR image is generated according to the echo amplitudes, lengths and azimuths of these 70 scattering centers.

[0027] Figure 7 Schematic diagram of the geometric relationship between the target and the radar provided in the embodiment of the present invention. Figure 7 The construction process of the simplified attribute scattering center model is explained.

[0028] like Figure 7 As shown in the figure, first establish the geometric relationship between the target and the radar. The dots in the target represent the scattering centers. is the origin of the coordinate system, for The unit vector in the direction is defined as , is the radar line of sight direction, which refers to the straight line direction from the radar transmitting antenna to the target. is an imaginary unit, Indicates the azimuth, the origin of the coordinate system The distance to the radar is , to the scattering center The distance is , then the scattering center on the target The path to the radar can be defined as ,in, , and Scattering centers The coordinate value of .

[0029] It should be understood that when the radar wavelength is much smaller than the target size, the target scattering field can be approximated by multiple scattered scattering centers, as shown below: ; In the formula, represents the target scattered field, Indicates the frequency of electromagnetic waves emitted by the radar. represents the propagation speed of electromagnetic waves, that is, the speed of light, Indicates the radar signal from the launch to the scattered center The time delay of the reflection back to the radar, , Represents the scattering center Frequency of electromagnetic waves and azimuth The dependence coefficient mainly depends on the physical and geometric properties of the scattering point itself. represents the complex amplitude, , Represents the scattering center The coordinate value of Indicates time, Is an imaginary unit.

[0030] Considering that the target is usually located in the far field, it can be ignored And the simplified target scattering field is obtained under given polarization conditions: ; Next, The frequency of electromagnetic waves emitted by radar and azimuth Modeling, specifically: based on geometric diffraction theory and physical optics analysis, using power function Frequency dependence factor of the scattering center Modeling, represents the imaginary unit, represents the center frequency of the radar; the azimuth angle dependence factor of the distributed scattering center is modeled using the sinc function, and the expression of the target scattering field is further simplified to obtain: ; In the formula, represents the attribute parameter set of the scattering center, Represents the scattering center The attribute parameter set, Represents the scattering center The echo, , , , Represent the scattering centers The echo amplitude, frequency dependence factor and length of Represents the scattering center Azimuth The azimuth angle dependence factor, Represents the scattering center The initial pointing angle.

[0031] Furthermore, the model is simplified again according to actual conditions to remove the influence of frequency dependence factors and radar azimuth domain.

[0032] On the one hand, the relative bandwidth of radar is usually small, resulting in The value of is close to 1, so the scattering center The frequency dependence factor Scattered field on target The influence of can be ignored, the scattering center Echo can be simplified to: .

[0033] On the other hand, due to the scattering center Azimuth The azimuth angle dependence factor is very small, so in the attribute scattering center model This term can be approximated as a constant and omitted.

[0034] Finally, the simplified attribute scattering center model can be obtained: ; .

[0035] It can be seen that in the simplified attribute scattering center model, the scattering center Attribute parameter set The number of parameters included is reduced from 7 to 5.

[0036] In the training process of the scattering feature extraction network, each round of training will select N The scattering characteristic parameters of the scattering center (i.e. N The echo amplitude, length, azimuth and coordinates of the scattering center in the original SAR image) are then sequentially converted into N Substitute the scattering characteristic parameters of the scattering center into , Pick , Pick , Pick , and then sum to get It should be understood that It is in matrix form, where each element corresponds to the value of each point in the reconstructed SAR image, so the reconstructed SAR image can be obtained through the above simplified attribute scattering center model.

[0037] Next, the SAR image scattering feature extraction method based on self-supervised learning provided by the present invention is further illustrated through simulation experiments.

[0038] Specifically, the public MSTAR (The Moving and Stationary Target Acquisition and Recognition) dataset is used in the simulation experiment to train and test the scattering feature extraction network. The MSTAR dataset is collected and released by Sandia National Laboratories in the United States. It contains a variety of targets and variants, different scenes and observation conditions, and various rotation angles and ground-grazing angles. The SAR images are acquired by an X-band synthetic aperture radar system. The resolution of the SAR image is 0.3m×0.3m, and the size ranges from 128×128 to 192×193. Since the SAR images at different azimuth angles are very different, several SAR images are collected in the range of 0°~360° in a circular manner with an azimuth angle of 1°~2°.

[0039] Figure 8 is a SAR image of 10 types of ground vehicle targets in the MSTAR dataset. The MSTAR dataset contains 10 types of ground vehicle targets, such as Figure 8As shown, they are: 2S1 (self-propelled howitzer), BMP2 (infantry fighting vehicle), BRDM2 (armored reconnaissance vehicle), BTR60 (armored transport vehicle), BTR70 (armored transport vehicle, model SN_C71), D7 (bulldozer), T62 (tank), T72 (tank, model SN_132), ZIL131 (cargo truck) and ZSU23 / 4 (self-propelled anti-aircraft gun).

[0040] The present invention extracts SAR image sets with a pitch angle of 17° and a pitch angle of 15° from the MSTAR data set, uses the SAR image with a pitch angle of 17° as a training set, and uses the SAR image with a pitch angle of 15° as a validation set, as shown in Table 1: Table 1

[0041] In order to retain the original information of the SAR image to the greatest extent, the size of the input original SAR image is 128×128. For SAR images with a size larger than 128×128, it is necessary to cut out the original SAR image with a size of 128×128 in the middle of the image.

[0042] In order to improve the stability and convergence speed of the scattering feature extraction network, the original SAR image can be normalized first. Since the original SAR image usually exists in complex form, it is necessary to splice the real part, imaginary part and amplitude value of the original SAR image and then input it into the scattering feature extraction network. In this way, the phase information in the original SAR image can be better utilized.

[0043] Fig. 9 It is a loss curve of the scattering feature extraction network on the training set and the data set, where the horizontal axis represents the number of training iterations, the vertical axis represents the loss value, train loss represents the training loss value, val loss represents the validation loss value, smooth train loss represents the smooth training loss value, and smooth val loss represents the smooth validation loss value. See Fig. 9 The loss value of the scattering feature extraction network decreases rapidly on the training set and the validation set, but gradually stabilizes and approaches 0 with the increase of training rounds, indicating that the above scattering feature extraction network has good convergence on the MSTAR dataset.

[0044] Furthermore, the existing scattering center parameter estimation method based on RELAX (relaxation algorithm) is selected for comparison with the present invention, and the results are shown in Table 2: Table 2

[0045] It can be seen from Table 2 that the average time for extracting the scattering features of each original SAR image by the scattering center parameter estimation method based on RELAX is 40 seconds, while the SAR image scattering feature extraction method based on self-supervised learning provided by the present invention only takes 0.073 seconds, which greatly improves the extraction efficiency of scattering features.

[0046] It should be noted that the scattering feature extraction results of SAR images exist in the form of parameter sets. In order to compare the extraction results intuitively and vividly, the simplified attribute scattering center model can be used to reconstruct the relevant parameters of the scattering features into SAR images.

[0047] Optionally, the accuracy of the present invention and the scattering center parameter estimation method based on RELAX are compared using two indicators: Structure Similarity Index Measure (SSIM) and Peak Signal to Noise Ratio (PSNR). SSIM is used to measure the structural similarity of two images, with a value range of 0 to 1. The closer the SSIM value is to 1, the higher the structural similarity of the two images. PSNR is used to indicate the peak signal-to-noise ratio between two images. The larger the PSNR value, the higher the similarity of the two images.

[0048] For the 10 types of ground vehicle targets in the validation set, the calculation results of the two indicators SSIM and PSNR are shown in Table 3: Table 3

[0049] It can be seen that after the reconstructed SAR images are generated based on the scattering features extracted by the two methods, the quality of the reconstructed SAR images is not much different.

[0050] Fig.10 This is a comparison chart of the reconstruction results of the original SAR image containing the type 2S1 ground vehicle target. Fig.11 This is a comparison chart of the reconstruction results of the original SAR image containing the type D7 ground vehicle target. Fig.12 This is a comparison chart of the reconstruction results of the original SAR image containing the T62 ground vehicle target. Fig.13 This is a comparison chart of the reconstruction results of the original SAR image containing the ZIL131 ground vehicle target. Fig.14 This is a comparison of the reconstruction results of the original SAR image containing the ZSU23 / 4 ground vehicle target. Figures 10-14In the figure, the left side shows the original SAR images of ground vehicle targets of types 2S1, D7, T62, ZIL131, and ZSU23 / 4, the middle side shows the reconstructed SAR image after scattering feature extraction based on the present invention, and the right side shows the reconstructed SAR image after scattering feature extraction using the scattering center parameter estimation method based on RELAX, which verifies the effectiveness of the present invention.

[0051] It can be seen from the above embodiments that the beneficial effects of the present invention are: (1) The present invention introduces a convolutional attention mechanism module into the scattering feature extraction network, which can fully extract the channel information and spatial information of the original SAR image, help the scattering feature extraction network to identify and enhance the scattering feature information of the input original SAR image, and capture long-range dependencies.

[0052] (2) The present invention uses different detection heads to predict different scattering characteristic parameters, thereby avoiding mutual interference between different scattering characteristic parameters in the same attribute parameter graph.

[0053] (3) In each round of training of the scattering feature extraction network, a reconstructed SAR image is generated using a pre-constructed simplified attribute scattering center model based on the predicted scattering feature parameters, namely the echo amplitude, length and azimuth of each scattering center in the original SAR image. The reconstructed SAR image is then incorporated into the calculation of the loss function value. The network parameters of the scattering feature extraction network are optimized based on the idea of ​​minimizing the difference between the reconstructed SAR image and the original SAR image, thereby achieving self-supervised learning without the need for pre-labeled labels. In addition, when constructing the simplified attribute scattering center model, the complex original attribute scattering center model is simplified, thereby improving the image reconstruction speed while ensuring the image reconstruction accuracy.

[0054] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0055] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the present application for which protection is sought, a person skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims.

[0056] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A SAR image scattering feature extraction method based on self-supervised learning, characterized in that: include: Get the original SAR image; The original SAR image is input into a scattering feature extraction network, so that the scattering feature extraction network performs the following steps: Using a convolution module to extract features from the original SAR image to obtain a first feature map; The first feature map is subjected to maximum pooling processing and then input into a residual module to obtain a second feature map; the residual module includes a plurality of residual units connected in sequence, each of which is used to generate a first sub-feature map according to an output of a residual unit at a previous level; Inputting the second feature map into a convolutional attention mechanism module to generate a third feature map including channel dimension feature information and spatial dimension feature information; The third feature map is input into a processing module, so that a plurality of sequentially connected processing units in the processing module generate a third sub-feature map according to two-way input data of the processing unit; wherein the two-way input data of the first-level processing unit in the processing module include: the third feature map and the first sub-feature map generated by the residual unit corresponding to the processing unit, the two-way input data of the last-level processing unit include: the third sub-feature map generated by the previous-level processing unit and the first feature map, and the two-way input data of the remaining processing units include: the third sub-feature map generated by the previous-level processing unit and the first sub-feature map generated by the residual unit corresponding to the processing unit; The third sub-feature map generated by the last-level processing unit is used as the fourth feature map, and is respectively input into the first detection head, the second detection head and the third detection head, and the first detection head, the second detection head and the third detection head perform parameter prediction based on their corresponding activation functions to obtain the scattering characteristics of the original SAR image.

2. The SAR image scattering feature extraction method based on self-supervised learning according to claim 1, characterized in that: The residual module includes a first residual unit, a second residual unit, a third residual unit and a fourth residual unit, and the processing module includes a first processing unit, a second processing unit, a third processing unit and a fourth processing unit; After the step of extracting features from the original SAR image using a convolution module to obtain a first feature map, the method further includes: inputting the first feature map into the first processing unit; The step of performing maximum pooling processing on the first feature map and then inputting the result into a residual module to obtain a second feature map comprises: The first residual unit generates a first sub-feature map based on the first feature map after the maximum pooling process. , and the first sub-feature map input into the second processing unit and the second residual unit respectively; The second residual unit is based on the first sub-feature map Generate the first sub-feature map , and the first sub-feature map input into the third processing unit and the third residual unit respectively; The third residual unit is based on the first sub-feature map Generate the first sub-feature map , and the first sub-feature map input into the fourth processing unit and the fourth residual unit respectively; The fourth residual unit is based on the first sub-feature map Generate the first sub-feature map , and the first sub-feature map As the second feature map, input the convolutional attention mechanism module.

3. The SAR image scattering feature extraction method based on self-supervised learning according to claim 2 is characterized in that: The convolutional attention mechanism module includes a channel attention unit and a spatial attention unit; The step of inputting the second feature map into a convolutional attention mechanism module to generate a third feature map including channel dimension feature information and spatial dimension feature information includes: The second feature map is input into the convolutional attention mechanism module, and the channel attention vector is extracted using the channel attention unit. , and based on the channel attention vector Weight the second feature map to generate a second sub-feature map ; The second sub-feature map Input the spatial attention unit and extract the spatial attention vector , and based on the spatial attention vector For the second sub-feature map Weighted, generate the third feature map.

4. The SAR image scattering feature extraction method based on self-supervised learning according to claim 3 is characterized in that: The second feature map is input into the convolutional attention mechanism module, and the channel attention vector is extracted using the channel attention unit. The steps include: Inputting the second feature map into the convolutional attention mechanism module, and performing maximum pooling processing and average pooling processing respectively; Inputting the second feature map after the maximum pooling process and the second feature map after the average pooling process into a shared multi-layer perceptron respectively, wherein the shared multi-layer perceptron is a two-layer fully connected network; The two outputs of the shared multilayer perceptron are added together and processed by the Sigmoid activation function to obtain the channel attention vector .

5. The SAR image scattering feature extraction method based on self-supervised learning according to claim 3 is characterized in that: The second sub-feature map Input the spatial attention unit and extract the spatial attention vector The steps include: The second sub-feature map The spatial attention unit is input, and the maximum pooling, average pooling and convolution processing are performed in sequence, and then the spatial attention vector is obtained by Sigmoid activation function. .

6. The SAR image scattering feature extraction method based on self-supervised learning according to claim 2, characterized in that: The step of inputting the third feature map into a processing module, so that a plurality of sequentially connected processing units in the processing module generate a third sub-feature map according to their own two-path input data, comprises: The third feature map is input into the processing module, and the third feature map is upsampled by the fourth processing unit and combined with the first sub-feature map After splicing, convolution processing is performed to generate the third sub-feature map ; The third processing unit processes the third sub-feature map Upsample and compare it with the first sub-feature map After splicing, convolution processing is performed to generate the third sub-feature map ; The second processing unit processes the third sub-feature map Upsample and compare it with the first sub-feature map After splicing, convolution processing is performed to generate the third sub-feature map ; The first processing unit processes the third sub-feature map Up-sample and concatenate with the first feature map and perform convolution to generate the third sub-feature map As the fourth characteristic map.

7. The SAR image scattering feature extraction method based on self-supervised learning according to claim 1, characterized in that: The step of using the third sub-feature map generated by the last-stage processing unit as the fourth feature map and inputting the fourth feature map into the first detection head, the second detection head and the third detection head respectively, and having the first detection head, the second detection head and the third detection head perform parameter prediction based on their corresponding activation functions to obtain the scattering characteristics of the original SAR image includes: Using the third sub-feature map generated by the last-stage processing unit as a fourth feature map, and inputting the fourth feature map into the first detection head, the second detection head, and the third detection head respectively; The first detection head generates a first attribute parameter map with the same scale as the original SAR image based on the ReLU activation function, and the value of each point in the first attribute parameter map represents the predicted echo amplitude of each scattering center in the original SAR image ; The second detection head generates a second attribute parameter map with the same scale as the original SAR image based on the improved Sigmoid activation function, and the value of each point in the second attribute parameter map represents the length of each scattering center in the original SAR image obtained by prediction. ; The third detection head generates a third attribute parameter map with the same scale as the original SAR image based on the improved Tanh activation function, and the value of each point in the third attribute parameter map represents the predicted initial pointing angle of each scattering center in the original SAR image .

8. The SAR image scattering feature extraction method based on self-supervised learning according to claim 1, characterized in that: The scattering feature extraction network is trained based on a mean square error loss function; Wherein, in each round of training of the scattering feature extraction network, the loss value of the mean square error loss function is: the mean value of the sum of squares of the differences between the pixel values ​​of the original SAR image and the reconstructed SAR image.

9. The SAR image scattering feature extraction method based on self-supervised learning according to claim 8, characterized in that: The reconstructed SAR image is generated according to the following steps: Obtain the echo amplitude, length, and azimuth of each scattering center in the original SAR image predicted during this round of training; Based on the echo amplitude of each scattering center in the original SAR image predicted during this round of training, the scattering centers in the original SAR image are arranged in descending order, and the top N scattering centers; Get the previous N The echo amplitude, length and azimuth of each scattering center are obtained and substituted into the pre-built simplified attribute scattering center model to generate a reconstructed SAR image.

10. The SAR image scattering feature extraction method based on self-supervised learning according to claim 9, characterized in that: The simplified attribute scattering center model is expressed as: ； ； In the formula, Represents the scattering center The attribute parameter set, , , , , Represent the scattering centers The echo amplitude, length and initial pointing angle, , Represents the scattering center The coordinate value of represents the Symle function, Indicates the frequency of electromagnetic waves emitted by the radar. is an imaginary unit, Indicates the transmission speed of electromagnetic waves. represents the azimuth, represents the target scattered field, Represents the scattering center The echo.

Images