ISAR image feature extraction method and application in non-cooperative spacecraft attitude estimation

Abstract

The invention discloses an ISAR image feature extraction method and application in non-cooperative spacecraft attitude estimation, and relates to the field of image processing and feature extraction, and the method comprises the steps: carrying out the preprocessing of an image in a data set; respectively carrying out shallow feature extraction and deep feature extraction on the preprocessed image based on transfer learning to obtain corresponding feature data Xd; wherein the deep feature extraction is realized by constructing a neural network based on a U structure, and an encoder part and a decoder part of the neural network both use a dense residual block DRB to complete multi-level feature extraction. According to the ISAR image feature extraction method and the application thereof in non-cooperative spacecraft attitude estimation, low-resolution features are extracted in combination with dense residual blocks of channel attention, the problem of sparse image features is relieved, feature loss caused by sub-sampling is compensated, and satellite imaging image space feature information is fully excavated.

Description

Technical Field

[0001] The present invention relates to image processing and feature extraction in the field of deep learning and computer vision technology. More specifically, the present invention relates to an ISAR image feature extraction method and its application in non-cooperative spacecraft attitude estimation. Background Art

[0002] Due to the lack of cooperative information, non-cooperative spacecraft in orbit cannot directly use sensors to collect attitude data, but can use Inverse Synthetic Aperture Radar (ISAR) image sequences to observe the attitude change information of space targets.

[0003] There are two traditional methods for spacecraft attitude estimation. One method is to estimate the structure, size and attitude of the spacecraft through three-dimensional reconstruction, which relies on obtaining the projection matrix of scattering points at different attitudes. Unfortunately, high-frequency radar measurements may cause phenomena such as "angular glint" and "cavity scattering", which makes feature correlation between ISAR images more challenging. Another method is template matching, which usually requires the acquisition of appropriate features such as radar cross section, HRRP and ISAR images. However, this method relies on prior information about the structure and size of the spacecraft, may be sensitive to noise, or is not suitable when the line of sight changes drastically. At the same time, deep learning-based methods have also initially adopted convolutional neural network modeling, but only use limited simulation data to train the model, and there are certain problems and limitations in practical applications.

[0004] Specifically, traditional methods rely heavily on prior conditions, so they face the following difficulties:

[0005] 1) High-frequency radar is prone to produce "angular glint" that affects matching;

[0006] 2) It depends on the target structure and size and is difficult to generalize.

[0007] When applying deep learning to existing spacecraft attitude estimation methods, due to the limitations of the number and quality of training samples, the model recognition effect is poor and it is difficult to meet the actual application needs. Therefore, deep learning methods face the following difficulties:

[0008] 1) The training samples are limited, the model is prone to overfitting, and the deep learning method is affected by the training data, making it difficult to reach industrial-grade application level.

[0009] 2) Design the network structure to optimize the feature extraction based on the low definition of satellite images, while ignoring the low resolution and blurred edge features of the images.

[0010] 3) Use convolutional neural networks to extract features, but ignore frequency domain information.

[0011] 4) The attention mechanism is used to fuse global information, but the network structure is not designed based on the imaging characteristics. Summary of the invention

[0012] An object of the present invention is to solve at least the above problems and / or disadvantages and to provide at least the advantages which will be described hereinafter.

[0013] In order to achieve these purposes and other advantages of the present invention, a method for extracting features from an ISAR image is provided, comprising:

[0014] S1, associating the ISAR image with the elevation and azimuth data therein to establish a data set;

[0015] S2, preprocess the images in the dataset;

[0016] S3, based on transfer learning, shallow feature extraction and deep feature extraction are performed on the preprocessed image to obtain the corresponding feature data X d ;

[0017] Among them, the deep feature extraction is achieved by constructing a neural network based on the U structure. The encoder and decoder parts of the neural network use dense residual blocks DRB to complete multi-level feature extraction, and each DRB reuses the previous layer features to fuse multi-scale information through residual connections.

[0018] Preferably, in S2, the preprocessing includes wavelet denoising and image enhancement using an adaptive histogram equalization method.

[0019] Preferably, in S3, the shallow feature X r Obtained by the following formula:

[0020] X r =ResNet(X+X e )

[0021] In the above formula, X is the original spacecraft ISAR image data, X e is the image data obtained after preprocessing, and ResNet(·) is the ResNet34 pre-training model.

[0022] Preferably, four dense residual blocks DRB are used for feature extraction in upsampling and downsampling of the neural network decoder part, and the last level of downsampling is connected to the first level of upsampling through an intermediate block, and the upsampling and downsampling of the same layer are connected through a jump connection;

[0023] Among them, each dense residual block DRB is composed of multiple residual blocks RB that are closely connected. Each RB uses depth-separable convolution to replace the original convolution operation, and each RB introduces a gating mechanism SG and a channel attention mechanism SCA, so that each RB in each DRB contains layer normalization LN, gating mechanism SG, and channel attention mechanism SCA;

[0024] Based on the given input feature X 1， And X1=X+Xe+Xr, then the output formula of RB RB(X1) is:

[0025] RB(X1)=X1+C1(SCA(SG(C3(C1(LN(X1))))))

[0026] The extracted feature of DRB DRB(X1) is:

[0027] DRB(X1)=RB(RB(X1)+X1)+RB(X1)+X1

[0028] In the above formula, C1 is a 1x1 convolution and C3 is a 3x3 depth convolution.

[0029] Preferably, the intermediate block is a dual-domain feature fusion;

[0030] The dual-domain feature fusion refers to the fusion of frequency domain feature extraction and spatial domain feature extraction.

[0031] Preferably, the frequency domain feature extraction is implemented based on the frequency domain gated network SGN, and the spatial domain feature extraction is implemented based on the multi-head self-attention block MHSAB. Then, for downsampling, the output feature XE of each layer DRB is i It can be expressed by the following formula:

[0032] XE0=None

[0033] XE1=DRB(X+X e +X r )

[0034] XE i =DRB(DOWN(XE1)+XE i-1

[0035] For the upper collection, the input feature XD of each layer of DRB i It can be expressed by the following formula:

[0036] XD4=XE4+UP(MHSAB(SGN(XE4)))

[0037] XD i =XE i +UP(XD i+1 )

[0038] Where i = 1, 2, 3, 4, UP(·) is the upsampling operation, DOWN(·) is the downsampling operation, SGN(·) is the frequency domain control network, MHSAB(·) is the multi-head self-attention block, DRB(·) is the extracted feature of DRB, X ris the shallow feature, X is the original image data, X e is the enhanced image data after preprocessing;

[0039] The deep feature X d It is expressed by the following formula:

[0040] X d =DRB(XD1).

[0041] An application of an ISAR image feature extraction method in non-cooperative spacecraft attitude estimation also includes:

[0042] S4, based on the feedforward network regression model for feature data X d Perform regression to obtain the final posture result;

[0043] Among them, the feedforward network regression model includes multiple neuron layers connected sequentially, and each neuron is only connected to the neurons in the previous layer. The transmission process is expressed by the following formula:

[0044]

[0045] In the above formula, is the predicted posture result value, W i is the weight vector.

[0046] Preferably, the difference between the predicted results and the actual results of different loss functions is compared, and the back propagation algorithm is used to update the network parameters so that the predicted value continues to be close to the actual value, and finally the feedforward network regression model training is realized.

[0047] The present invention has at least the following beneficial effects:

[0048] Firstly, the present invention combines wavelet denoising with a CLAHE method to propose a comprehensive signal-to-noise ratio optimization method, which makes up for the defect that the traditional method is prone to generate noise.

[0049] Secondly, the present invention effectively alleviates the problem that the deep learning model is limited by the number and quality of training samples through the transfer learning method, and improves the recognition accuracy.

[0050] Thirdly, the present invention proposes to extract low-resolution features by using dense residual blocks combined with channel attention, which alleviates the problem of sparse image features, compensates for the feature loss caused by subsampling, and fully mines the spatial feature information of satellite imaging images.

[0051] Fourthly, the present invention combines the frequency domain gating network with the multi-head self-attention mechanism, and introduces frequency domain information to support posture recognition in this task for the first time, which improves the feature expression ability compared with the existing methods.

[0052] Fifth, the network architecture design of the present invention is innovative, integrating multidisciplinary knowledge and expanding the application prospects of deep learning in this field.

[0053] Other advantages, objectives and features of the present invention will be embodied in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a schematic diagram of the process of the ISAR image feature extraction method of the present invention;

[0055] Figure 2 It is a residual block RB structure diagram of the present invention combined with channel attention (SCA);

[0056] Figure 3 This is a dense residual block DRB structure diagram composed of multiple RBs used in the present invention;

[0057] Figure 4 It is a structural diagram of the middle block of the present invention;

[0058] Figure 5 is the original spacecraft ISAR image data;

[0059] Figure 6 This is a schematic diagram of the effect of directly performing histogram equalization processing on the original image or the image converted into a grayscale image in advance;

[0060] Figure 7 It is a schematic diagram of the wavelet denoising process of the present invention;

[0061] Figure 8 This is a schematic diagram of the process of processing an image using wavelet denoising in the present invention;

[0062] Fig. 9 Schematic diagram of the angle error curve before and after the attitude estimation network model is improved;

[0063] Fig.10 Schematic diagram of the PFU-Net architecture for ISAR imaging spacecraft attitude estimation of the present invention. DETAILED DESCRIPTION

[0064] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0065] The ISAR image feature extraction method of the present invention is as follows: Figure 1 As shown, it mainly includes the following parts:

[0066] ① Combining wavelet denoising with adaptive histogram equalization (CLAHE) method, a comprehensive signal-to-noise ratio optimization method is proposed to make up for the defect of traditional methods that are prone to noise.

[0067] ② Use transfer learning to make up for the lack of training data.

[0068] ③ Design a dense residual block combined with channel attention to extract low-resolution features for ISAR imaging, achieve adaptive feature refinement, and reduce feature loss in the downsampling process.

[0069] ④ Combine the frequency domain gating network with multi-head self-attention to introduce additional frequency domain information and obtain more global feature information.

[0070] Specifically, the ISAR image feature extraction method and its application in non-cooperative spacecraft attitude estimation mainly include:

[0071] 1. Dataset preparation

[0072] Data acquisition is a prerequisite, and the purpose is to provide a high-quality data set for subsequent image feature extraction and attitude regression model training. It is necessary to obtain ISAR images and elevation and azimuth data and associate them to build a complete data set. This data set will become the basis of our research.

[0073] The pathlib library is used to load the paths of all images in the folder, and the pytesseract-OCR technology is used to extract the text. The angle information and image path are integrated and associated based on character matching. The OCR tool successfully identifies and extracts the angle information in each spacecraft image, and matches these extracted angle parameters with the original image path to form a data set based on ISAR images.

[0074] A dataset of 4100 on-orbit ISAR images from a specific spacecraft was used. The dataset was divided into a training set containing 80% of the data and a test set containing 20% ​​of the data.

[0075] 2. Image preprocessing

[0076] In the image preprocessing stage, image enhancement will be performed to improve image quality. Compare various histogram equalization processing methods and select a method that is more suitable for spacecraft ISAR images to reduce image noise. Use CLAHE (Contrast Limited Adaptive Histogram Equalization) technology to enhance the contrast of the image and reduce the impact of noise. This will help improve the usability of the image, but it is not directly applicable to the image. Figure 5Although histogram equalization can improve the contrast of the original image or the image converted into grayscale in advance, it will increase the noise, such as Figure 6 As shown, the noise is obviously larger than the original image:

[0077] Normally, between noise reduction and enhancement, noise reduction is given priority. This is because noise will affect the subsequent enhancement effect. This step combines wavelet noise reduction with the adaptive histogram equalization method (CLAHE) to enhance contrast and clarity and detail information, so a comprehensive signal-to-noise ratio optimization method is proposed to make up for the defect that traditional methods are prone to noise. In practical applications, if the image noise is large, wavelet transform can be performed first. This is because noise will affect the equalization processing effect of adaptive histogram equalization (CLAHE), and image enhancement processing is performed after noise reduction. Therefore, the image is subjected to wavelet transform in advance, and the process of wavelet noise reduction is as follows: Figure 7 As shown in the figure, the variables in the noise reduction process are low frequency, horizontal high frequency, vertical high frequency, and diagonal high frequency. The image after noise reduction is as follows: Figure 8 As shown:

[0078] After the wavelet transform, the image is enhanced to improve the contrast. By applying the adaptive histogram equalization method, the structural similarity of the image is improved and the contrast is increased. The contrast is effectively improved and the noise is reduced.

[0079] 3. Design a network model for attitude estimation of spacecraft ISAR images

[0080] Data-driven methods usually require a large amount of data to train the model so that it can capture the characteristics of the data and establish a mapping relationship between input and output. However, the amount of spacecraft ISAR image data is small and the data features are sparse, which makes it difficult to train the model. To this end, transfer learning is introduced to use the pre-trained model to extract the shallow features of the image and initialize the shallow parameters of the model to speed up the training process.

[0081] This step mainly designs a deep neural network model based on the improved U-Net (PFU-Net) network architecture, applies the neural network to extract features from the images obtained by the inverse synthetic aperture radar (ISAR), and then applies it to the attitude estimation of non-cooperative satellite targets in the later stage. It mainly includes:

[0082] (1) Shallow feature extraction: First, the pre-trained model ResNet is used to extract shallow features from the processed image. This step aims to capture the primary features of the image and provide useful information for subsequent deep feature extraction. At the same time, in order to retain the original ISAR image features, the original spacecraft ISAR image features are added to the enhanced features and input into the ResNet34 pre-trained model to obtain the shallow feature X. r, let X be the original spacecraft ISAR image data, Xe be the enhanced image data after processing and enhancement, Xr and Xd represent the feature maps extracted from the shallow layer and deep layer after model processing, respectively. The specific expression of Xr is as follows:

[0083] X r =ResNet(X+X e )

[0084] (2) Deep feature extraction: A U-structured neural network (PFU-Net) is constructed to extract deep features. The encoder-decoder part of the network will use DRB (Dense Residual Blocks) blocks, which help to complete multi-level feature extraction. Each DRB block reuses the previous layer features, fuses multi-scale information through residual connections, reduces feature loss, and captures more complex image features.

[0085] In the process of using ordinary two-dimensional convolution to extract input data features, traditional neural networks will simultaneously extract the spatial features and channel features of the input data, so there is redundancy in the features. In the classic residual network structure, the convolutions used are all ordinary two-dimensional convolutions. Therefore, as the depth of the network model increases, the number of parameters and the amount of calculation of the network model also increase dramatically. However, the non-cooperative spacecraft ISAR images obtained by the present invention belong to a typical small sample data set. If an overly large model is used, overfitting is more likely to occur. The design uses a deep separable convolution instead of an ordinary convolution in the dense residual block of the residual network structure, which can not only obtain deeper feature information of the data to a certain extent, but also separate and extract the spatial features and channel features of the input data, reduce redundant features, and achieve better feature extraction effects for the model.

[0086] Specifically, the neural network structure based on the U structure mainly includes: the coordinated use of dense residual blocks and dual-domain feature fusion. The structural features of each part are explained below:

[0087] 1) Dense residual block

[0088] In order to effectively solve the problem of reducing system complexity and resource consumption, the present invention proposes a dense residual block combined with channel attention (SCA). Each dense residual block (DRB) is composed of multiple residual blocks (RBs) tightly connected, such as Figure 3 Each residual block RB uses a depth-separable convolution to replace the original convolution operation, and introduces a simple gating mechanism SG in each residual block (RB). The SCA channel attention mechanism is introduced after the gating mechanism, as shown in Figure 2As shown, each RB includes layer normalization (Layer Normalization, LN for short), deep separable convolution (Deep Convolution, DCnv for short), gating mechanism (Simple Gate, SG for short), channel attention mechanism (Simplelified ChannelAttention, SCA for short) and residual connection. The residual connection of the present invention is a dense connection, which directly uses 1x1 convolution operation to realize information exchange between channels, promotes the fusion of global information and local information, reduces feature loss in the transmission process, and reduces resource consumption and system complexity.

[0089] Given an input feature X 1， And X1=X+Xe+Xr, then the output formula of RB RB(X1) is:

[0090] RB(X1)=X1+C1(SCA(SG(C3(C1(LN(X1))))))

[0091] In the above formula, SCA(·) is the output formula of SCA, SG(·) is the output formula of SG, and LN((·) is the output formula of LN(;

[0092] Assuming that the input gate feature is represented by F, the output formula of SG SG(F) is:

[0093] SG(F)=F f1 ×F f2

[0094] RB(X1)=X1+C1(SCA(SG(C3(C1(LN(X1))))))

[0095] F f1和 F f2 Represents the different partial features of the input feature F in the channel interval, that is, decomposing the feature F into two parts F f1 and F f2 , just like the two terminals of a switch. The features on the two terminals interact through multiplication, which has the effect of selective amplification and compression. f1 The value of a certain position in the middle is large, F f2 When the corresponding position in is also large, their product will be larger, and the feature is selectively amplified. On the contrary, when they are inconsistent, the product tends to 0, and the corresponding feature is selectively compressed. Just like gating, selectively amplifying the signal is conducive to extracting key information.

[0096] The feature DRB(X1) is extracted by dense residual block (DRB):

[0097] DRB(X1)=RB(RB(X1)+X1)+RB(X1)+X1

[0098] Among them, C1 is 1x1 convolution, C3 is 3x3 depth convolution, LN is layer normalization, SG is gating, and SCA is channel attention.

[0099] 2) Dual-domain feature fusion (frequency domain + spatial domain)

[0100] The Spectral Gating Network (SGN) is a neural network module proposed in recent years. Its key innovation is to convert input features into the frequency domain for processing and introduce a learnable gating mechanism, aiming to enhance the ability of neural networks to model and express information in the frequency domain.

[0101] like Figure 4 As shown, the design uses a frequency domain gating network (SGN) combined with a multi-head self-attention block (MHSAB) as a connector (middle block) between the encoder and decoder, and designs a spectral layer to capture different frequency components of the image to understand the local frequency. It consists of a fast Fourier transform (FFT) layer, followed by a weighted gating layer, and then a fast inverse Fourier transform (IFFT) layer. The spectral layer uses FFT to convert the physical space to the spectral space. A learnable weight parameter is used to determine the weight of each frequency component to obtain the lines and edges of the image. The spectral space is restored to the physical space using the inverse fast Fourier transform (IFFT). Finally, the final feature map data is processed using global self-attention. Convolution reduces the sampling space, and attention can effectively process smaller resolutions, thereby improving performance. In this way, the network can learn data patterns in the frequency domain, introduce additional features in the frequency domain, and enhance feature expression capabilities. From the last layer of the encoder output XE4, the frequency domain feature Xs extraction formula is as follows:

[0102] X s =IFFT2(FFT2(XE4)⊙W)

[0103] Where W is weighted gating, FFT is Fourier transform, and IFFT is fast Fourier transform. In this design, this part of the process can be expressed as Xs=SGN(XE4), where XE4 represents the features of the downsampled 4th layer encoder output, and Xs is the frequency domain network extraction result.

[0104] The multi-head self-attention mechanism was first used by the Transformer model. First, multiple linear mappings are performed on the features output by the frequency domain gating network to obtain multiple sets of query matrices Q, key matrices K, and value matrices V. The mapping results are input into the scaled dot product attention. The result obtained each time is called a head (in this scheme, the input features are first linearly mapped multiple times to obtain multiple sets of query matrices Q, key matrices K, and value matrices V. Each set of Q, K, and V calculates the scaled dot product attention to obtain a set of attention heads. Q, K, and V are query matrices, key matrices, and value matrices, respectively).

[0105] In the present invention, starting from the deep features Xs output by the frequency domain feature network (SGN), MHSAB first applies a 1x1 convolution to generate the query matrix Q∈R H×W×d , the bond matrix K∈R H×W×d The sum matrix V∈R H×W×d , as shown below:

[0106] Q=C1(X S )

[0107] K=C1(X S )

[0108] V=C1(X S )

[0109] Among them, X S It is the deep feature output by the frequency domain feature network, which is a three-dimensional tensor. H is the size of the feature tensor in the height direction, and W is the size of the feature tensor in the width direction. C is the channel dimension of the feature tensor, R represents the total number of elements of the feature tensor on the H×W plane, which is equal to H×W, d is the dimension of the new feature space after the multi-head self-attention layer transformation, and N is the batch dimension. Many samples are calculated at the same time. It should be noted that the frequency domain feature network converts the last layer feature map X4 of the encoder into the frequency domain, adds gating, and then converts it back to the spatial domain. Therefore, the output of the frequency domain feature network is a new deep feature containing time-frequency features.

[0110] In order to effectively associate information between position-aware objects, the attention mechanism considers both content information and the relative distances between features at different positions. The encoding representing the highly relative distance position is Ph∈R H×1×d , indicating the encoding of the width relative to the distance position is Pw∈R 1×W×d The logical calculation is: QK T +QP T , where QK T Representing content-content interaction, QP T Represents content-location interaction.

[0111] The self-attention mechanism calculates the attention matrix as:

[0112] Attention(Q,K,P,V)=SoftMax(QK T +QP T )V

[0113] Among them, Q represents the query matrix (Query), K represents the key matrix (Key), V represents the value matrix (Value), T represents the matrix transpose, and P is the position information.

[0114] In this design, this part of the process can be expressed as Xm=MHSAB(Xs), where Xs is the frequency domain network extraction result and Xm is the feature extraction result of multi-head attention.

[0115] By incorporating a self-attention mechanism into the model, we can compute the correlations between different positions in the sequence and assign weights to each position. This allows us to globally focus on the sequence and dynamically model long-distance dependencies. In addition, it helps reduce parameter coupling, thereby enhancing the expressiveness and performance of the model. Finally, we reshape the attention matrix back to R H×W×d The resulting output is then added to XE4 and passed to the decoder. By using this approach, we can gather more information and improve the overall performance of the model.

[0116] In a specific application, the enhanced feature Xe and the shallow feature Xr are added to the input feature X, and they are input into the dense residual connection. Formally, let XE i is the output feature of the i-th layer (i=1,2,3,4) of the encoder part. The PFU-Net architecture for ISAR imaging spacecraft attitude estimation is shown in Fig.10 As shown. Then at each level i of the encoder part, the downsampled input feature XE i It is expressed as:

[0117] XE0=None

[0118] XE1=DRB(X+X e +X r )

[0119] XE i =DRB(DOWN(XE1)+XE i-1

[0120] At each level i in the decoder part, the input feature XD is upsampled i It is expressed as:

[0121] XD4=XE4+UP(MHSAB(SGN(XE4)))

[0122] XD i =XE i +UP(XDi+1 )

[0123] In the above formula, UP(·) is the upsampling operation, DOWN(·) is the downsampling operation, SGN(.) is the frequency domain control network, MHSAB(·) is the multi-head self-attention block, and DRB(·) is the dense residual block.

[0124] Finally, the deep feature X is obtained through the dense residual network d It is expressed as:

[0125] X d =DRR(XD1)

[0126] It can be seen that the output of the first encoder block is passed to all subsequent encoder blocks. This method helps to reduce feature loss and improve the overall performance of the model.

[0127] (3) Feedforward Network

[0128] The deep feature Xd obtained by the improved dense residual network is regressed through the feedforward network to obtain the final posture result. The feedforward network consists of neuron layers connected in a sequential manner. Each neuron is only connected to the neurons in the previous layer. It receives the output of the previous layer and provides output to the next layer. At the same time, it provides its own output to the next layer. The process is as follows:

[0129]

[0130] In order to train the model PFU-Net proposed in this paper, three different loss functions are used for model training. They are L1 loss (also known as mean absolute error or MAE), L2 loss (also known as mean square error or MSE) and Huber loss. Through these loss functions, the model is back-propagated and parameter updated to minimize the difference between the predicted results and the true results. It should be noted that the back-propagation algorithm belongs to the existing deep learning technology. This scheme compares the performance of different loss functions under this task, selects the loss function that can better fit the data distribution and local sparse samples, calculates the difference between the predicted results and the true results in the loss function, and uses the back-propagation algorithm to update the network parameters so that the predicted value continues to be close to the actual value, and finally realizes model training.

[0131] The L1 loss calculation formula is:

[0132]

[0133] Where n is the number of samples, y i represents the value of the i-th true sample, Represents the model prediction value.

[0134] MAE measures the mean absolute error between the predicted values ​​and the true values.

[0135] The L2 loss calculation formula is:

[0136]

[0137] Unlike MAE, MSE uses the sum of squares of errors, which can pay more attention to the impact of large errors. i , The meaning of is the same as MAE. By calculating the error values ​​of these two loss functions on the training set and the validation set, researchers can evaluate the generalization ability of the model under different loss functions.

[0138] In spacecraft ISAR imaging, due to the limitations of ISAR equipment, the training data contains individual outliers. By minimizing the impact of a single anomaly, the model inadvertently harms its performance on other common examples, resulting in a decrease in overall performance. To overcome this limitation, this paper uses the Huber loss function, which is less sensitive to outliers in the data, as follows:

[0139]

[0140] This function essentially computes the absolute error, but when the error is small, it switches to the mean squared error. The hyperparameter δ controls the sensitivity of this error.

[0141] Specifically, for a given spacecraft ISAR image X, we first apply wavelet transform and adaptive histogram equalization to enhance the image quality and obtain Xe. Next, the original image X and the enhanced image Xe are extracted from the shallow feature map Xr through the pre-trained model. Then these shallow features Xr, Xe and X are passed through a 4-level encoder, decoder and an intermediate block that combines frequency domain and spatial domain (frequency domain gating + multi-head self-attention) to obtain the deep feature Xd. Finally, we apply convolution and feedforward networks to the deep feature Xd and generate the pose result.

[0142] Example 1:

[0143] The application of Tiangong-1 Inverse Synthetic Aperture Radar (ISAR) images is used to illustrate the application, using 4800 on-orbit ISAR imaging datasets from a specific spacecraft. The dataset is divided into a training set containing 80% of the data and a test set containing 20% ​​of the data. The amount of spacecraft ISAR image data is small and the data features are sparse, which makes it difficult to train the model. In order to speed up the training, transfer learning is introduced to extract shallow features and initialize the shallow parameters of the pre-trained model. ISAR imaging technology can only detect spacecraft within a limited range, resulting in the attitude angles of all images being significantly concentrated between 0° and 35° and between 5° and 10°.

[0144] The effect of image preprocessing is as follows Figure 5 As shown in the figure, in order to extract the implicit quantitative parameters in the image, optical character recognition (OCR) technology is used to detect and process each sample image. The OCR tool successfully identifies and extracts the angle information of the spacecraft image. The extracted angle parameters are matched with the original image path to form a data set based on ISAR images. The distribution of the number of samples at different angles in the data set is statistically sorted, and the representativeness of the data set in the subsequent attitude estimation network training is preliminarily verified.

[0145] Use Fig.10 The network structure shown performs transfer learning, dense residual connection, and then the feedforward network regression model gives the final prediction result.

[0146] In order to test the improved model's performance in predicting the attitude of spacecraft ISAR images, we used our valuable data set for training and testing. The angle error curves before and after the model improvement are shown in Figure 2. Fig. 9 As shown in the figure, the horizontal axis is the number of training times and the vertical axis is the loss value during the training process. Fig. 9 It can be seen that before the model was improved, the initial error value was about 4.2°, close to 5°. After the model was improved, the initial error was about 2, which was reduced by about 52.4%. In the initial stage of training, including the later stage, the error value before improvement showed an upward trend many times, indicating that overfitting occurred. However, the error value designed by the present invention continued to show a steady downward trend, and no overfitting occurred. At the end of the training, the error value before the model was improved was about 2.6, and the error value after the model was improved was about 0.5, which decreased by 80.77%. It can be seen that the error is significantly reduced after the model is improved, and it is optimized through the fitting phenomenon, which can be verified that the model designed by the present invention is more detailed in feature extraction on non-cooperative on-orbit spacecraft ISAR images, and the prediction results are more accurate.

[0147] The above solution is only an illustration of a preferred embodiment, but is not limited thereto. When implementing the present invention, appropriate replacement and / or modification can be performed according to user needs.

[0148] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily realized. Therefore, without departing from the general concept defined by the claims and equivalent scope, the present invention is not limited to the specific details and the illustrations shown and described here.

Claims

1. A method for extracting features from an ISAR image, characterized in that: include: S1, associating the ISAR image with the elevation and azimuth data therein to establish a data set; S2, preprocess the images in the dataset; S3, based on transfer learning, shallow feature extraction and deep feature extraction are performed on the preprocessed image to obtain the corresponding feature data X d ; Among them, the deep feature extraction is achieved by constructing a neural network based on the U structure. The encoder and decoder parts of the neural network use dense residual blocks DRB to complete multi-level feature extraction, and each DRB reuses the previous layer features to fuse multi-scale information through residual connections.

2. The ISAR image feature extraction method according to claim 1, characterized in that: In S2, the preprocessing includes wavelet denoising and image enhancement using an adaptive histogram equalization method.

3. The ISAR image feature extraction method according to claim 1, characterized in that: In S3, the shallow feature X r Obtained by the following formula: X r =ResNet(X+X e ) In the above formula, X is the original spacecraft ISAR image data, X e is the image data obtained after preprocessing, and ResNet(·) is the ResNet34 pre-training model.

4. The ISAR image feature extraction method according to claim 1, wherein: Four dense residual blocks (DRBs) are used for feature extraction in the upsampling and downsampling of the neural network decoder. The last level of downsampling is connected to the first level of upsampling through an intermediate block, and the upsampling and downsampling in the same layer are connected through skip connections. Among them, each dense residual block DRB is composed of multiple residual blocks RB that are closely connected. Each RB uses depth-separable convolution to replace the original convolution operation, and each RB introduces a gating mechanism SG and a channel attention mechanism SCA, so that each RB in each DRB contains layer normalization LN, gating mechanism SG, and channel attention mechanism SCA; Based on the given input feature X 1， And X1=X+Xe+Xr, then the output formula of RB RB(X1) is: RB(X1)=X1+C1(SCA(SG(C3(C1(LN(X1)))))) The extracted feature of DRB DRB(X1) is: DRB(X1)=RB(RB(X1)+X1)+RB(X1)+X1 In the above formula, C1 is a 1x1 convolution and C3 is a 3x3 depth convolution.

5. The ISAR image feature extraction method according to claim 4, characterized in that: The middle block is a dual-domain feature fusion; The dual-domain feature fusion refers to the fusion of frequency domain feature extraction and spatial domain feature extraction.

6. The ISAR image feature extraction method according to claim 5, characterized in that: The frequency domain feature extraction is implemented based on the frequency domain gated network SGN, and the spatial domain feature extraction is implemented based on the multi-head self-attention block MHSAB. For downsampling, the output feature XE of each layer DRB i It can be expressed by the following formula: XE0=None XE1=DRB(X+X e +X r ) CAR i =DRB(DOWN(XE1)+XE i-1 For the upper collection, the input feature XD of each layer of DRB i It can be expressed by the following formula: XD4=XE4+UP(MHSAB(SGN(XE4))) XD i =CAR i +UP(XD i+1 ) Where i = 1, 2, 3, 4, UP(·) is the upsampling operation, DOWN(·) is the downsampling operation, SGN(·) is the frequency domain control network, MHSAB(·) is the multi-head self-attention block, DRB(·) is the extracted feature of DRB, X r is a shallow feature, X is the input original image data, X e is the enhanced image data after preprocessing; The deep feature X d It is expressed by the following formula: X d =DRB(XD1)。 7. An application of the ISAR image feature extraction method as claimed in claim 1 in non-cooperative spacecraft attitude estimation, characterized in that: Also includes: S4, based on the feedforward network regression model for feature data X d Perform regression to obtain the final posture result; Among them, the feedforward network regression model includes multiple neuron layers connected sequentially, and each neuron is only connected to the neurons in the previous layer. The transmission process is expressed by the following formula: In the above formula, is the predicted posture result value, W i is the weight vector.

8. Application of the ISAR image feature extraction method as claimed in claim 7 in non-cooperative spacecraft attitude estimation, characterized in that By comparing the differences between the predicted results and the actual results caused by different loss functions, the back-propagation algorithm is used to update the network parameters so that the predicted values ​​continue to be close to the actual values, and finally the feedforward network regression model training is realized.

Images