SAR image self-focusing method and system based on error coefficient optimization

Abstract

The invention discloses an SAR (Synthetic Aperture Radar) image self-focusing method and system based on error coefficient optimization. The method comprises the following steps: training a feature extraction module by using a defocused SAR image sample to establish a mapping relation between a defocused SAR image and a polynomial error coefficient; for a to-be-processed defocused SAR image # imgabs0 #, the defocused SAR image # imgabs1 # passes through a trained feature extraction module to obtain a polynomial error coefficient, the polynomial of a phase error vector is converted into an error matrix # imgabs2 #, the defocused SAR image # imgabs3 # is subjected to azimuth fast Fourier transform # imgabs4 # and then multiplied by the error matrix # imgabs5 #, and the defocused SAR image # imgabs3 # is obtained. And performing inverse fast Fourier transform # imgabs6 # on a result obtained by multiplication in the azimuth direction to obtain a focused SAR image # imgabs7 #. The SAR self-focusing method is small in manual intervention, high in calculation efficiency and better in interpretability, and the quality and effect of SAR image self-focusing are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing high-resolution imaging, and in particular relates to a SAR image self-focusing method and system based on error coefficient optimization. Background Art

[0002] SAR images are images acquired through synthetic aperture radar (SAR) technology. Unlike optical imaging, the essence of SAR imaging is the result of the accumulation of SAR echo signals in azimuth. It can use the movement of the radar platform to increase the effective observation aperture, thereby achieving a higher resolution in azimuth than the physical antenna size. This synthetic aperture radar technology transmits and receives electromagnetic waves at different locations, accumulates data from multiple observations, and then compresses it in azimuth to form a high-resolution image. Figure 1 As shown in the figure, when the carrier aircraft is in actual motion, the actual motion trajectory of the platform deviates from the ideal trajectory due to the influence of airflow disturbance and the irregular motion of the platform itself, which leads to phase error during azimuth compression in the SAR imaging process. The generation of phase error will affect the quality of SAR imaging and make the imaging effect blurred. In order to obtain high-quality SAR images, it is very necessary to eliminate the motion error generated during the movement of the carrier aircraft. At present, the SAR image motion compensation methods mainly include coarse compensation methods based on motion data and image self-focusing methods based on echo data. Among them, the compensation method based on motion data is more dependent on the measurement accuracy of the inertial navigation system (INS) and the global positioning system (GPS), and this method can only compensate for large motion errors. To obtain high-quality SAR images, it is necessary to estimate the Doppler parameters from the echo data and perform error compensation to achieve precise focusing of the SAR image.

[0003] like Figure 2As shown in the figure, SAR image autofocusing methods can be divided into model-driven methods and data-driven methods. Model-driven SAR image autofocusing methods can be divided into phase-based and amplitude-based estimation methods. Among them, the phase-based estimation methods include phase gradient method (PGA) and phase difference method (PDA). The performance of this type of method is affected by the window size, the compensation ability for high-frequency phase errors is weak, and strong scattering points are required in the image. The amplitude-based estimation methods mainly include two types of methods: visual misalignment (MDA) and metric-based optimization. Among them, the error order upper limit of the subview-based SAR autofocusing method is limited by the size of the subaperture, and the estimation of the phase error is not accurate enough; while the metric optimization method is based on the classic image minimum entropy and maximum contrast theory, which can well compensate for the phase error of any order frequency, but its optimization process requires repeated iterative calculations, which is time-consuming, and for SAR images with different degrees of defocus, the model parameter settings are not the same, and the manual intervention required is large. In addition, the application of its algorithm model does not take into account the environmental interference factors that may appear in real scenes, resulting in poor application effect.

[0004] In recent years, with the rapid development of big data theory and computing power, data-driven technologies represented by deep learning have gradually emerged. This type of technology does not rely too much on prior theories, but is based on optimization techniques such as gradient backpropagation, resorting to obtaining models from large amounts of data training. It has a wider range of applications and has stronger performance in models obtained by computing power support optimization. Through nested neural layers (convolutional layers based on matrix operations for images) and nonlinear activation functions, deep neural networks can establish a mapping model from input images to optimization indicators. By setting appropriate loss functions and training the model, adaptive learning of algorithm parameters can be achieved, thus avoiding the tedious steps of setting prior parameters of traditional algorithms and repeated trial and error iterative operations. For example, Mason et al. proposed a synthetic aperture radar self-focusing imaging method based on recursive neural networks, applying deep learning technology to the field of SAR imaging for the first time. Wei Pu proposed a sparse autoencoder network (SAE-Net), in which the encoder is designed for SAR imaging through the ADMM algorithm, and the decoder is used for reconstruction mapping from imaging results to SAR echoes. The network training is guided by reconstruction loss and entropy loss to generate focused SAR images. Liu Zhi proposed a non-iterative autofocus method based on deep learning and minimum entropy criterion, using entropy as the loss function to guide the network to automatically learn the autofocus rules, thereby achieving automatic focusing of SAR images. However, for the current SAR autofocus methods based on deep learning, most methods directly establish a mapping relationship between the input image and the focused image, and achieve the effect of SAR image focusing through network training. This type of method cannot explain the reason why SAR images are defocused, is not interpretable, and cannot calculate the specific phase error. Summary of the invention

[0005] Technical problem to be solved by the present invention: In view of the above-mentioned problems in the prior art, a SAR image self-focusing method and system based on error coefficient optimization are provided. The present invention aims to realize a SAR self-focusing method with less manual intervention, high computational efficiency and better interpretability, so as to improve the quality and effect of SAR image self-focusing.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: A SAR image self-focusing method based on error coefficient optimization comprises the following steps: S1, using the defocused SAR image samples to train a feature extraction module to establish defocused SAR images, polynomial error coefficients ~ The mapping relationship between them, and the polynomial error coefficient ~ and the phase error vector of the defocused SAR image The polynomial is: ,in is the order, is the azimuth frequency vector; S2, for the defocused SAR image to be processed , the defocused SAR image The polynomial error coefficients are obtained by the trained feature extraction module ~ , the polynomial error coefficient ~ Using the phase error vector The polynomial is converted into an error matrix , the defocused SAR image Fast Fourier Transform of azimuth After the error matrix Multiply, and then the result of the multiplication is subjected to the inverse fast Fourier transform in the azimuth direction Get focused SAR image .

[0007] Optionally, in step S2, the polynomial error coefficient ~ Convert to error matrix The function expression is: , In the above formula, is the diagonalization operation, is an imaginary unit, is the phase error vector, and the phase error vector at this time is the polynomial error coefficient ~ and defocused SAR images The azimuth frequency vector Using the phase error vector The polynomial calculation is obtained.

[0008] Optionally, the polynomial error coefficient in step S1 ~ The feature extraction module trained in step S1 includes a global feature branch and a local feature branch in parallel, and extracts the global features of the defocused SAR image through the global feature branch and maps them to the low-order error coefficients in the polynomial error coefficients, and extracts the local features of the defocused SAR image through the local feature branch and maps them to the high-order error coefficients in the polynomial error coefficients.

[0009] Optionally, extracting the global features of the defocused SAR image through the global feature branch and mapping them into low-order error coefficients in the polynomial error coefficients comprises: inputting the defocused SAR image into the Restormer network to extract the global features f b , the global feature f b The characteristics f b The low-order error coefficient in the polynomial error coefficient is obtained by passing through a global pooling layer, a flattening layer, a random dropout layer and a fully connected layer in sequence, where the flattening layer is composed of a fully connected layer and a LeakyReLU activation layer.

[0010] Optionally, extracting the local features of the defocused SAR image through the local feature branch and mapping them into high-order error coefficients in the polynomial error coefficients includes: inputting the defocused SAR image into the CNN network to extract the local features f a , the local features f a After resizing, it passes through a global pooling layer, a flattening layer, a random dropout layer, and a fully connected layer in turn to obtain the high-order error coefficient in the polynomial error coefficient, where the flattening layer is composed of a fully connected layer and a LeakyReLU activation layer.

[0011] Optionally, in step S1, a feature extraction module is trained using the defocused SAR image samples to establish the defocused SAR image and the polynomial error coefficients. ~ When the mapping relationship between them is established, it includes training the feature extraction module using the training data set and updating the network parameters of the feature extraction module using a preset loss function, and the function expression of the preset loss function is: , In the above formula, is the loss function, and are the sizes of the defocused SAR image in azimuth and range respectively, is the SAR amplitude image The element in row i and column j in is the power of the SAR echo signal, .

[0012] Optionally, when the feature extraction module is trained using the training data set and the network parameters of the feature extraction module are updated using a preset loss function, the optimizer used is an AdamW optimizer, and the function expression of the network parameters of the feature extraction module updated using the preset loss function by the AdamW optimizer is: , , , In the above formula, and are the momentum at the tth and t-1th iterations, respectively, and are the exponential decay rates of the gradient and the square of the gradient, respectively. is the gradient of the loss function at the tth iteration, and are the variances at the t-th and t-1-th iterations, represents element-wise multiplication, and are the network parameters of the feature extraction module RCM at the t+1th and tth iterations, respectively. is the learning rate, is a numerical stability term used to prevent the denominator from being zero, is the weight decay term.

[0013] In addition, the present invention also provides a SAR image self-focusing system based on error coefficient optimization, comprising a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the SAR image self-focusing method based on error coefficient optimization.

[0014] In addition, the present invention also provides a computer-readable storage medium, in which a computer program or instruction is stored. The computer program or instruction is programmed or configured to execute the SAR image self-focusing method based on error coefficient optimization through a processor.

[0015] In addition, the present invention also provides a computer program product, including a computer program or an instruction, wherein the computer program or the instruction is programmed or configured to execute the SAR image autofocusing method based on error coefficient optimization through a processor.

[0016] Compared with the prior art, the present invention has the following advantages: the method of the present invention comprises using defocused SAR image samples to train a feature extraction module to establish a mapping relationship between the defocused SAR image and the polynomial error coefficient; for the defocused SAR image to be processed, , the defocused SAR image The polynomial error coefficients are obtained by the trained feature extraction module, and the polynomial of the phase error vector is converted into an error matrix , the defocused SAR image Fast Fourier Transform of azimuth After the error matrix Multiply, and then the result of the multiplication is subjected to the inverse fast Fourier transform in the azimuth direction Get focused SAR image The present invention realizes a SAR self-focusing method with less manual intervention, high computational efficiency and better interpretability, which can improve the quality and effect of SAR image self-focusing. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of an airborne SAR imaging model in the prior art.

[0018] Figure 2 The present invention classifies the SAR image autofocusing methods in the prior art.

[0019] Figure 3 Schematic diagram of the basic flow of the method of the embodiment of the present invention.

[0020] Figure 4 Schematic diagram of the principle of the method of the embodiment of the present invention.

[0021] Figure 5 Schematic diagram of the network structure of the feature extraction module of an embodiment of the present invention.

[0022] Figure 6 These are three original SAR defocused images selected in the embodiment of the present invention.

[0023] Figure 7 SAR focused image comparison diagram obtained by different methods in the embodiments of the present invention.

[0024] Figure 8 It is a comparison chart of qualitative indicator results of ablation experiments in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The specific implementation modes of the present invention will be further described in detail below with reference to the accompanying drawings.

[0026] like Figure 3 and Figure 4 As shown, the SAR image self-focusing method based on error coefficient optimization in this embodiment includes the following steps: S1, using the defocused SAR image samples to train a feature extraction module to establish defocused SAR images, polynomial error coefficients ~ The mapping relationship between them, and the polynomial error coefficient ~ and the phase error vector of the defocused SAR image The polynomial is: ,in is the order, is the azimuth frequency vector; S2, for the defocused SAR image to be processed , the defocused SAR image The polynomial error coefficients are obtained by the trained feature extraction module ~ , the polynomial error coefficient ~ Using the phase error vector The polynomial is converted into an error matrix , the defocused SAR image Fast Fourier Transform of azimuth After the error matrix Multiply, and then the result of the multiplication is subjected to the inverse fast Fourier transform in the azimuth direction Get focused SAR image .

[0027] The SAR image autofocusing method based on MEA (minimum entropy autofocusing) estimates the phase error in the azimuth direction during SAR imaging by iterative optimization and compensates the original image. The phase error in the range direction is generally small and can be ignored. Therefore, the estimation and compensation of the azimuth phase error is usually carried out in the range Doppler domain. For the convenience of research, it is assumed that is the complex matrix of the SAR defocused image in the range Doppler domain, then the azimuth phase error compensation problem of the image can be defined as follows: , In the above formula, is the compensated SAR image The pixel point at the azimuth a and range r in the SAR image, where a and r represent the azimuth and range indexes of the SAR image, respectively. and Represents SAR images The number of points in the azimuth and distance directions, yes The corresponding phase error vector, For SAR images Direction in and distance The pixel position, is an imaginary unit; the defocused SAR image Fast Fourier Transform of azimuth After the error matrix Multiply, and then the result of the multiplication is subjected to the inverse fast Fourier transform in the azimuth direction Get focused SAR image It can be expressed as: , In the above formula, The phase error vector The converted matrix is for The diagonalization operation, and are the azimuth fast Fourier transform operation and the azimuth inverse fast Fourier transform operation respectively. From the above formula, we can see that for any input defocused SAR image X , if the corresponding phase error matrix can be calculated in this embodiment The traditional MEA method takes the phase error vector Modeling can be performed based on actual physical factors. Modeled as a polynomial error or a sinusoidal error. In this embodiment, this embodiment takes a polynomial type error as an example, and its phase error vector can be modeled as follows: , In the above formula, is the error vector coefficient, is the order, is the azimuth frequency vector, which can generally be normalized to [-0.5, 0.5] or [-1, 1]; the constant term and the first-order term are ignored in the above polynomial, mainly because the constant term does not affect the entropy of the image, and the first-order term only affects the Doppler center frequency, resulting in an azimuth shift of the image, but does not affect the image focusing quality. In this embodiment, the seventh-order polynomial error is taken as an example, and its expression is as follows: .

[0028] In step S2 of this embodiment, the polynomial error coefficient ~ Convert to error matrix The function expression is: , In the above formula, is the diagonalization operation, is an imaginary unit, is the phase error vector, and the phase error vector at this time is the polynomial error coefficient ~ and defocused SAR images The azimuth frequency vector Using the phase error vector The polynomial calculation is obtained.

[0029] The polynomial error coefficient in step S1 of this embodiment ~ It is composed of two parts: low-order error coefficient and high-order error coefficient, which can be divided as needed. The feature extraction module trained in step S1 includes parallel global feature branches and local feature branches, and extracts the global features of the defocused SAR image through the global feature branch and maps it to the low-order error coefficient in the polynomial error coefficient, and extracts the local features of the defocused SAR image through the local feature branch and maps it to the high-order error coefficient in the polynomial error coefficient. The traditional metric-based SAR image self-focusing method continuously optimizes the SAR image by establishing the relationship between the phase error and the SAR image metric index to achieve the self-focusing function. This method is relatively inefficient and has certain blindness. On the one hand, the network does not establish a relationship with the metric value from the image itself, and the network training has a large blindness; on the other hand, directly taking the phase error as the parameter to be optimized, the function mapping relationship is relatively complex and the model convergence speed is slow. Based on the shortcomings of the traditional method, this embodiment proposes a new SAR image self-focusing method. The method of this embodiment is to predict the phase error by predicting the error coefficient from the image itself, and this embodiment also divides the error coefficient into high-order coefficients and low-order coefficients. This method is more friendly to the model prediction of the SAR image phase error. like Figure 4 This is an end-to-end network that starts from the original SAR image and generates the final SAR focused image. The network can be divided into three parts: feature extraction, parameter estimation, and self-focus compensation.

[0030] For the seventh-order polynomial phase error expression established in the previous text, this embodiment also needs to predict the unknown error coefficients therein to obtain a set of coefficient values ​​that optimize the output image index. In this embodiment, the phase error of the SAR image can be divided into low-order phase error and high-order phase error, which affect the main lobe and side lobe of the SAR echo signal respectively. In this embodiment, the following assumption is made: the low-order phase error mainly affects the basic shape and structure of the SAR image, causing large-scale blurring of the image; while the high-order phase error will affect the details and texture of the image, resulting in blurring or distortion of the details. In view of this assumption, this embodiment regards the 2nd, 3rd, and 4th order polynomial errors in the method proposed in this embodiment as low-order phase errors, and the 5th, 6th, and 7th order polynomial errors as high-order phase errors. In order to obtain the predicted values ​​of the low-order phase error coefficient and the high-order phase error coefficient value, the method of this embodiment is that for a SAR image, the global features and local features of the image are first extracted respectively, and then the extracted features are mapped to low-order error coefficients and high-order error coefficients respectively. The deep learning network based on CNN and Transformer has great advantages in extracting local features and global features of SAR images. In this embodiment, a variant of Transformer, Restormer, is used as a global feature extraction method, mainly because Transformer needs to calculate the relationship between each position of the self-attention layer and other positions during feature extraction, resulting in high computational cost and slow processing speed for high-resolution SAR images. Unlike traditional Transformer, Restormer adopts multi-head attention and feedforward network, and replaces the traditional self-attention layer applied across spatial dimensions in Transformer by applying self-attention mechanism across channels, thereby greatly reducing the complexity of model calculation.

[0031] like Figure 5 As shown, in this embodiment, the global features of the defocused SAR image are extracted through the global feature branch and mapped to the low-order error coefficients in the polynomial error coefficients, including: inputting the defocused SAR image into the Restormer network to extract the global features f b , the global feature f b The characteristics f bThe low-order error coefficients in the polynomial error coefficients are obtained by sequentially passing through a global pooling layer, a flattening layer, a random drop layer, and a fully connected layer, wherein the flattening layer is composed of a fully connected layer and a LeakyReLU activation layer. The Restormer network is an existing network that combines the characteristics of CNN and Transformer. It mainly includes an Embedding layer, a Transformer encoder, an attention module, an upsampling layer, and an activation function layer, and its output is the same size as the input. In order to keep consistent with the output feature size of the CNN feature extraction network, this embodiment compresses the features extracted by the Restormer network in the distance direction, so that the number of channels becomes 8 times the original, and then passes through a GAP layer so that the output result size is also N×16×1×256.

[0032] like Figure 5 As shown, in this embodiment, the local features of the defocused SAR image are extracted by the local feature branch and mapped into high-order error coefficients in the polynomial error coefficient, including: inputting the defocused SAR image into the CNN network to extract the local features f a , the local features f a After resizing, it passes through a global pooling layer, a flattening layer, a random dropout layer, and a fully connected layer in turn to obtain the high-order error coefficients in the polynomial error coefficients, where the flattening layer is composed of a fully connected layer and a LeakyReLU activation layer. In this embodiment, let is the input sample of the defocused SAR image, where N is the number of SAR image samples and the number of channels is 2. n a and n r Respectively represent the size of the SAR image in azimuth and distance. In this embodiment n a and n r All are set to 256. For the CNN feature extraction network, its network structure mainly consists of an input layer, a convolutional layer, an instance normalization (IN), a LeakyReLU activation layer, and a global average pooling layer (GAP). The network configuration and the output size of each layer are shown in Table 1.

[0033] Table 1 CNN feature extraction network layer input and output configuration table

[0034] After the defocused SAR image is subjected to feature extraction by the CNN network and the Restormer network, a feature of uniform scale is obtained. Then the extracted features need to be converted into error coefficients. For the parameter estimation part, its network structure mainly consists of a flattening layer, a random dropout layer, and a fully connected (FC) layer. The flattening layer is mainly used to flatten the result obtained by the global average pooling layer into a column vector to form a 16×256-dimensional output result. The random dropout layer (p=0.5) is a regularization method, which means that neurons are randomly dropped with a probability of 0.5 to prevent network training from overfitting. Finally, after passing through a fully connected (FC) layer of (16×256)×3 size, the network output size is a column vector of size N×3 (N is the number of samples). Then the two groups of N×3 vectors obtained by parameter estimation of the features extracted by the CNN network and the Restormer network can be expressed as high-order error coefficients and low-order error coefficients, respectively. The specific network layer configuration is shown in Table 2.

[0035] Table 2 Parameter estimation network layer input and output configuration table

[0036] For any defocused SAR image with a size of 256×256 and a channel number of 2, such as Figure 5 The figure shows a schematic diagram of obtaining the estimated values ​​of the polynomial coefficients after the feature extraction and parameter estimation process.

[0037] After the defocused SAR complex image is processed through feature extraction and parameter estimation network, the polynomial error coefficient estimation value of the image can be output. Then, the obtained coefficient estimation value needs to be used to compensate the defocused SAR image to obtain the result of single compensation, and the optimal network structure is obtained by setting the loss function and continuously optimizing. The global and local features extracted by the Restormer branch and the CNN branch are respectively mapped to the low-order terms of the polynomial error coefficient after network processing ( ) and higher-order terms ( ), the phase error vector of the SAR image and the SAR image after self-focus compensation can be calculated respectively. In step S1 of this embodiment, a feature extraction module is trained using defocused SAR image samples to establish a defocused SAR image, a polynomial error coefficient ~ When the mapping relationship between them is established, it includes training the feature extraction module using the training data set and updating the network parameters of the feature extraction module using a preset loss function, and the function expression of the preset loss function is: , In the above formula, is the loss function, and are the sizes of the defocused SAR image in azimuth and range respectively, is the SAR amplitude image The element in row i and column j in is the power of the SAR echo signal, .

[0038] According to the calculated loss function, the polynomial error coefficient is adjusted and updated by the optimizer, and finally the optimal solution of the phase error vector is obtained. Common optimizers used for optimizing parameters include SGD optimizer, LBFGS optimizer, Adam optimizer, etc. The AdamW optimizer used in this embodiment adds a weight decay term on the basis of the Adam optimizer to better improve the generalization ability of the model. Specifically, when this embodiment uses the training data set to train the feature extraction module and uses the preset loss function to update the network parameters of the feature extraction module, the optimizer used is the AdamW optimizer, and the optimizer used is the AdamW optimizer. The function expression of the network parameters of the feature extraction module updated by the preset loss function is: , , , In the above formula, and are the momentum at the tth and t-1th iterations, respectively, and are the exponential decay rate of the gradient and the exponential decay rate of the square of the gradient (usually set to 0.9, 0.999), is the gradient of the loss function at the tth iteration, and are the variances at the t-th and t-1-th iterations, represents element-wise multiplication, and are the network parameters of the feature extraction module RCM at the t+1th and tth iterations, respectively. is the learning rate, is a numerical stability term used to prevent the denominator from being zero (usually set to 10 -8 ), is the weight decay term (can be set to 10 -3 ).

[0039] In order to verify the SAR image self-focusing method based on error coefficient optimization in this embodiment, the data used in the experiment comes from the Advanced Land Observing Satellite (ALOS) satellite developed by the Earth Observation Research Center of the Japan Aerospace Exploration Agency, which is equipped with a phased array L-band synthetic aperture radar (PALSAR). PALSAR has three working modes: high-resolution mode, scanning mode and multi-polarization mode. In this embodiment, we use SAR images acquired in high-resolution mode, which have high resolution and high signal-to-noise ratio in the range direction, with a resolution better than 2m; in azimuth, the SAR image in this mode has an azimuth resolution of about 5m and a ground resolution of about 7m. For ease of processing, we randomly cropped the original entire SAR image and cropped it into small images of 256×256 size, containing a total of 11,000 images. In this experiment, we distribute it according to the ratio of training set: test set = 10:1. The relevant parameters of ALOS PALSAR are shown in Table 3.

[0040] Table 3 Parameters of ALOS PALSAR in high-resolution mode

[0041] For spaceborne SAR, there are many factors that cause phase errors, such as off-track motion, altitude changes, atmospheric effects, and parameter setting errors during the imaging process. These errors will reduce the quality of imaging and cause the SAR image to be defocused. This process can be modeled as the result of convolution of the undefocused image with a blur filter. For the training set and test set of this embodiment, we simulate this process by randomly generating polynomial errors in the range Doppler domain for the undefocused image, thereby obtaining a defocused SAR image. This is also a common practice in the field of SAR imaging.

[0042] In order to verify whether the method proposed in this embodiment is effective for SAR image self-focusing, this experiment proposes to use entropy and contrast as evaluation indicators of experimental results. These indicators are briefly introduced below.

[0043] a. Entropy. In SAR self-focusing imaging, image entropy can be used to measure the focusing quality of SAR images, and the smaller the entropy value, the clearer the focused SAR image and the better the quality. Its definition is as follows: , In the above formula, is the image entropy, and the other parameters are the same as those of the loss function L.

[0044] b. Contrast. Similar to entropy, contrast is also an indicator to measure the clarity of SAR images. The higher the contrast value, the clearer the focused SAR image is and the better the image quality is. There is no uniform definition of contrast. The common ones are defined as the ratio of the target energy mean square error to the target energy mean and the ratio of the target energy mean to the square of the target amplitude mean. In this embodiment, we use the former as the contrast definition method, and its expression is as follows: , In the above formula, is the contrast, is the amplitude representation of the input SAR defocused image X, It means to find the mathematical expectation.

[0045] In this example, all method implementations were trained and tested on a high-performance workstation equipped with an Nvidia Tesla A100 GPU (80 GB memory) and an AMD Ryzen Threadripper PRO 5995WX 64-core CPU. The deep learning framework is PyTorch, using CUDA version 11.7. In the experiment, both the training set and the test set were cropped into 256×256 image data, the number of training rounds was set to 100 rounds, and the batch size was set to 8. The initial learning rate was set to 10 -3 , decreasing by 50% every 20 rounds.

[0046] In order to verify the effectiveness of the method proposed in this embodiment, the experiment compares the self-focusing results obtained by the method proposed in this embodiment with the results obtained by the traditional MEA algorithm and the self-focusing method based on Bagging-ECELMs. At present, there are still few studies on SAR self-focusing methods based on deep learning. The self-focusing method based on Bagging-ECELMs has a better self-focusing effect and is very comparable. The following will compare them from two aspects: qualitative indicators and quantitative indicators.

[0047] 1) Qualitative indicator results.

[0048] For the sake of convenience, the experiment selected three pictures of different scenes from the test set as representatives for display. Figure 6 The figure shows three original SAR defocused images in the test set. It can be seen from the images that there are some strong scattering points in the image, and the defocus phenomenon is more obvious at these strong scattering points. Some detailed textures and contours become blurred, which seriously affects the overall imaging performance of the SAR image. The purpose of the SAR image self-focusing method is to compress the SAR echo signal in the azimuth direction to eliminate the defocus phenomenon in the image and make the image clear.

[0049] like Figure 7 The figure shows the comparison results of the focused images obtained by the MEA algorithm, the Bagging-ECELMs algorithm and the method of this embodiment for the three original SAR defocused images in the test set. It is not difficult to see from the figure that the greater the scattering intensity, the more likely it is to defocus, and the more serious the defocusing degree. The red boxes in the figure are some contours or textures in each image, and the yellow circles show the imaging information of the strong scattering points. The strong scattering points in the first row and third column are elongated in azimuth due to severe defocusing, and the yellow boxes are local enlarged images of the strong scattering point area in the yellow circle. From the comparison results in the figure, the SAR images obtained by the three self-focusing methods are clearer than the original SAR defocused images, and the resolution is improved. In more detail, in the parts marked with red boxes in scenes 1 and 2, the SAR images obtained by the method of this embodiment have clearer contours and clearer textures; and in scene 3, from the local enlarged image marked by the yellow box, it can be seen that the strong scattering points in the SAR image obtained by the method of this embodiment are more concentrated in azimuth, and the focusing effect is better. In summary, the method proposed in this embodiment can realize the self-focusing function of SAR images, and has better focusing effect on features such as long strip contours, textures, and strong scattering points compared with the traditional MEA method and the Bagging-ECELMs method based on deep learning.

[0050] In order to more objectively verify the effectiveness of the method of this embodiment, the experimental results obtained are verified based on the entropy and contrast indicators proposed in the previous article. Table 4 shows the quantitative index results of the original SAR defocused image in the test set, the SAR focused image obtained by the MEA method, the Bagging-ECELMs method, and the method of this embodiment, where the bold values ​​are the optimal values ​​of each indicator.

[0051] Table 4 Comparison of quantitative index results of SAR focused images obtained by different methods

[0052] From the index results listed in Table 4, the method of this embodiment performs best in both entropy and contrast indicators, and the entropy index is reduced by 0.188 relative to the original image test set, and the contrast index is improved by 1.483 relative to the original image test set. The experimental results show that the method of this embodiment can indeed improve the imaging quality of SAR images, and the method of this embodiment is superior to the MEA method and the Bagging-ECELMs method in terms of self-focusing effect.

[0053] In order to better verify the hypothesis proposed in this embodiment, the method of this embodiment sets up an ablation experiment, mainly to compare the method of using the RCM module to predict the high-order error coefficients and the low-order error coefficients and the method of only using the CNN network to predict the error coefficients, which method is better for the SAR image self-focusing effect. In terms of network settings, only the Restormer feature extraction network needs to be replaced with the CNN feature extraction network, while other settings remain unchanged. According to the operations in the comparative experiment, we selected three SAR images of different scenes from the test set as a display, and the obtained SAR focused images are shown as follows: Figure 8 As shown. As can be seen from the figure, using only the CNN network to predict the coefficients can also achieve the effect of focusing on the SAR image. However, in more detail, for the white texture and black outline in the red box in scene 1 and scene 2, the SAR image texture obtained by our method is more obvious and the outline is clearer; and for the local area of ​​the scattering points circled by the yellow circle in scene 3, the focusing effect of the two for strong scattering points is not much different, and both can achieve a good focusing effect. This also explains our hypothesis: low-order phase errors mainly affect the basic shape and structure of the SAR image, causing large-scale blurring of the image; while high-order phase errors affect the details and texture of the image, resulting in blurring or distortion of the details. It is reasonable.

[0054] In addition, from the perspective of quantitative indicators, Table 5 records the quantitative index comparison of the SAR focused image obtained by using only the CNN network prediction error coefficient method and the SAR focused image obtained by the RCM module prediction error coefficient method proposed in this paper.

[0055] Table 5 Quantitative index results of ablation experiment

[0056] Referring to Figure 5, it can be seen that, numerically, the method of this embodiment is optimal in terms of the two given indicators, contrast and entropy, and compared with the method using only the CNN network (Only CNN), the entropy is 0.38 lower and the contrast is 0.435 higher, which further illustrates the effectiveness and rationality of the method of this embodiment.

[0057] In summary, the method of this embodiment is based on the traditional MEA SAR image self-focusing idea, and proposes a new SAR image self-focusing method. This method designs a feature extraction module from the SAR image itself to predict the phase error coefficient, and uses the powerful feature extraction capabilities of neural networks such as CNN and Restormer to extract the detail features and global features of the SAR image respectively; secondly, this embodiment models the azimuth phase error as a polynomial phase error, and divides its coefficients into high-order error coefficients and low-order error coefficients. The detail features and global features extracted according to RCM are mapped to the high-order error coefficients and low-order error coefficients of the polynomial error respectively; then, the phase error of the input SAR image is calculated according to the predicted polynomial error coefficient, and the original SAR image is compensated; finally, the network can output the best SAR image focusing result after continuous optimization training by the optimizer. In experimental verification, the method of this embodiment is compared with the traditional MEA method and the deep learning method based on Bagging-ECELMs, and an ablation experiment is also designed. The experimental results can reflect the effectiveness and advancement of the method of this embodiment.

[0058] In addition, this embodiment also provides a SAR image autofocusing system based on error coefficient optimization, including a microprocessor and a memory connected to each other, and the microprocessor is programmed or configured to execute the SAR image autofocusing method based on error coefficient optimization.

[0059] In addition, this embodiment also provides a computer-readable storage medium, in which a computer program or instruction is stored. The computer program or instruction is programmed or configured to execute the SAR image autofocusing method based on error coefficient optimization through a processor.

[0060] In addition, this embodiment also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute the SAR image self-focusing method based on error coefficient optimization through a processor.

[0061] Those skilled in the art should understand that the technical solutions provided by the embodiments of the present application may be in the form of methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the process Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0062] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A SAR image self-focusing method based on error coefficient optimization, characterized in that: The steps include: S1, using the defocused SAR image samples to train a feature extraction module to establish defocused SAR images, polynomial error coefficients ~ The mapping relationship between them, and the polynomial error coefficient ~ and the phase error vector of the defocused SAR image The polynomial is: ,in is the order, is the azimuth frequency vector; S2, for the defocused SAR image to be processed , the defocused SAR image The polynomial error coefficients are obtained by the trained feature extraction module ~ , the polynomial error coefficient ~ Using the phase error vector The polynomial is converted into an error matrix , the defocused SAR image Fast Fourier Transform of azimuth After the error matrix Multiply, and then the result of the multiplication is subjected to the inverse fast Fourier transform in the azimuth direction Get focused SAR image .

2. The SAR image autofocusing method based on error coefficient optimization according to claim 1, characterized in that: In step S2, the polynomial error coefficient ~ Convert to error matrix The function expression is: ， In the above formula, is the diagonalization operation, is an imaginary unit, is the phase error vector, and the phase error vector at this time is the polynomial error coefficient ~ and defocused SAR images The azimuth frequency vector Using the phase error vector The polynomial calculation is obtained.

3. The SAR image autofocusing method based on error coefficient optimization according to claim 1, characterized in that: Polynomial error coefficients in step S1 ~ The feature extraction module trained in step S1 includes a global feature branch and a local feature branch in parallel, and extracts the global features of the defocused SAR image through the global feature branch and maps them to the low-order error coefficients in the polynomial error coefficients, and extracts the local features of the defocused SAR image through the local feature branch and maps them to the high-order error coefficients in the polynomial error coefficients.

4. The SAR image autofocusing method based on error coefficient optimization according to claim 3, characterized in that: The method of extracting the global features of the defocused SAR image through the global feature branch and mapping them into low-order error coefficients in the polynomial error coefficients includes: inputting the defocused SAR image into the Restormer network to extract the global features f b , the global feature f b The characteristics f b The low-order error coefficient in the polynomial error coefficient is obtained by passing through a global pooling layer, a flattening layer, a random dropout layer and a fully connected layer in sequence, where the flattening layer is composed of a fully connected layer and a LeakyReLU activation layer.

5. The SAR image autofocusing method based on error coefficient optimization according to claim 3, characterized in that: The method of extracting local features of the defocused SAR image through the local feature branch and mapping them into high-order error coefficients in the polynomial error coefficients includes: inputting the defocused SAR image into the CNN network to extract local features f a , the local features f a After resizing, it passes through a global pooling layer, a flattening layer, a random dropout layer, and a fully connected layer in turn to obtain the high-order error coefficient in the polynomial error coefficient, where the flattening layer is composed of a fully connected layer and a LeakyReLU activation layer.

6. The SAR image autofocusing method based on error coefficient optimization according to claim 1, characterized in that: In step S1, a feature extraction module is trained using the defocused SAR image samples to establish the defocused SAR image and the polynomial error coefficients. ~ When the mapping relationship between them is established, it includes training the feature extraction module using the training data set and updating the network parameters of the feature extraction module using a preset loss function, and the function expression of the preset loss function is: ， In the above formula, is the loss function, and are the sizes of the defocused SAR image in azimuth and range respectively, is the SAR amplitude image The element in row i and column j in is the power of the SAR echo signal, .

7. The SAR image autofocusing method based on error coefficient optimization according to claim 6, characterized in that: When the feature extraction module is trained using the training data set and the network parameters of the feature extraction module are updated using the preset loss function, the optimizer used is the AdamW optimizer, and the function expression of the network parameters of the feature extraction module updated using the preset loss function by the AdamW optimizer is: ， ， ， In the above formula, and are the momentum at the tth and t-1th iterations, respectively, and are the exponential decay rates of the gradient and the square of the gradient, respectively. is the gradient of the loss function at the tth iteration, and are the variances at the t-th and t-1-th iterations, represents element-wise multiplication, and are the network parameters of the feature extraction module RCM at the t+1th and tth iterations, respectively. is the learning rate, is a numerical stability term used to prevent the denominator from being zero, is the weight decay term.

8. A SAR image autofocusing system based on error coefficient optimization, comprising a microprocessor and a memory connected to each other, characterized in that: The microprocessor is programmed or configured to execute the SAR image autofocusing method based on error coefficient optimization as recited in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program or instruction stored therein, characterized in that: The computer program or instruction is programmed or configured to execute the SAR image autofocusing method based on error coefficient optimization according to any one of claims 1 to 7 through a processor.

10. A computer program product comprising a computer program or instructions, characterized in that The computer program or instruction is programmed or configured to execute the SAR image autofocusing method based on error coefficient optimization according to any one of claims 1 to 7 through a processor.

Images