SAR scene target identification method and device based on lightweight neural network
Through the SAR scene object recognition method based on lightweight neural network, a multi-input lightweight neural network system is built using the amplitude, phase and wavelet domain information of the SAR signal, which solves the problem of low SAR image recognition performance in the prior art, and realizes the The effects of high accuracy and rapid training inject new vitality into the field of SAR image recognition.
Patent Information
- Application Number
- CN202411007982.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-15
AI Technical Summary
The existing neural network model performs poorly in SAR image recognition, mainly because it ignores the unique signal nature and internal correlation of SAR images, and does not fully consider the physical principles of optical imaging and SAR imaging, resulting in complex scenarios Recognition performance degraded.
A SAR scene object recognition method based on lightweight neural network was designed. By taking the amplitude, phase and wavelet domain information of the SAR signal as input, it constructs including input preprocessing, basic feature extraction, core feature extraction, deep feature extraction, and space. A system for mapping and classification output, combined with multi-input lightweight neural network, streamlines the network structure to improve identification performance.
It significantly improves the accuracy of SAR image recognition, ensures more than 93% of the recognition accuracy, and achieves a 1.1-fold increase in training speed while maintaining the same hyperparameter settings, making the method more practical in emergency scenarios.
Smart Images

Figure CN118968146A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of computer neural network and signal processing, and in particular to a SAR scene target recognition method and device based on a lightweight neural network. Background Art
[0002] The continuous progress and improvement of deep learning theory has greatly promoted the development of neural networks in the field of image recognition, especially in optical image recognition. With the help of deep learning models, such as convolutional neural networks (CNN), researchers have been able to achieve extremely high accuracy and efficiency in image classification and target detection. However, when trying to apply these techniques to SAR image recognition, the challenges faced increase significantly.
[0003] SAR images, as a unique type of remote sensing data, have very different imaging principles, data features, and information presentation methods from optical images. SAR captures surface information by actively emitting microwaves and receiving their reflected signals. Its images contain rich texture, phase, and multi-polarization features, but the extraction of this information is often accompanied by problems such as speckle noise and geometric distortion. These characteristics make the recognition task of SAR images more complex than that of optical images.
[0004] Although deep learning has achieved remarkable results in the field of optical image recognition, it is not reliable to directly use existing neural network models for SAR image recognition. This is because there are essential differences between optical images and SAR images in data distribution and feature expression, which leads to the fact that models that perform well in optical image recognition may not be able to effectively extract key features when processing SAR images, thus affecting the recognition effect. Therefore, we need to design and optimize a special neural network model based on the characteristics of SAR images.
[0005] At present, in the field of SAR image recognition, although deep learning methods based on convolutional neural networks have shown certain results on specific SAR target data sets, their performance has dropped significantly when faced with complex scenes. The reason behind this is that these methods often directly apply the grayscale images of SAR images to neural networks customized for optical images, thereby ignoring the unique signal nature of SAR images and their internal correlations. In addition, these methods tend to directly use the network parameters pre-trained on optical images during training, but fail to fully consider the fundamental differences in the physical principles of optical imaging and SAR imaging, especially the rich phase information contained in SAR images. Therefore, when faced with real SAR scenes and target recognition tasks, the performance of these methods is greatly reduced.
[0006] Therefore, how to design an innovative SAR scene target recognition method has become a problem to be solved. Summary of the invention
[0007] To this end, the present invention provides a SAR scene target recognition method and device based on a lightweight neural network. The method takes the amplitude, phase and wavelet domain information of the SAR signal as input, can deeply explore and fully utilize the diverse features of SAR images, thereby significantly improving the recognition performance and injecting new vitality into the field of SAR image recognition.
[0008] In order to achieve the above object, the present invention provides the following technical solution: a SAR scene target recognition method based on a lightweight neural network, comprising:
[0009] Determine the performance index of process optimization according to the specified production process; determine the process range that affects the performance index according to the performance index;
[0010] Constructing a SAR scene target recognition system based on a lightweight neural network, the SAR scene target recognition system includes an input preprocessing submodule, a basic feature extraction submodule, a core feature extraction submodule, a deep feature extraction submodule, a spatial mapping submodule and a classification output submodule;
[0011] Constructing a training data set, and training the SAR scene target recognition system through the training data set;
[0012] Training a setting module of the SAR scene target recognition system through the training data set, and saving weight parameters of the setting module after training;
[0013] Loading the weight parameters trained by the setting module into the SAR scene target recognition system to obtain the trained SAR scene target recognition system;
[0014] The SAR scene target signal is input into the trained SAR scene target recognition system, and the trained SAR scene target recognition system is used to perform recognition to obtain a final recognition result.
[0015] As a preferred solution of a SAR scene target recognition method based on a lightweight neural network, the input preprocessing submodule is used to annotate the training data set and divide the training data set into a training set and a test set; by randomly scattering the samples, the amplitude, phase and wavelet domain information of the samples are extracted;
[0016] The basic feature extraction submodule extracts edge information of SAR scene targets and understands the shape and contour of the targets through the edge information.
[0017] As a preferred solution of a SAR scene target recognition method based on a lightweight neural network, the core feature extraction submodule is used to extract local connection information of SAR scene targets, and reveal the local structure and relationship inside the target through the local connection information;
[0018] The deep feature extraction submodule is used to extract deep internal structure information, and capture high-level properties of SAR scene targets through the internal structure information.
[0019] As a preferred solution of a SAR scene target recognition method based on a lightweight neural network, the spatial mapping submodule uses a multi-layer perceptron network and a spatial mapping strategy to perform weighted mapping on the features received from the deep feature extraction submodule to obtain a category data value, and sends the category data value to the classification output submodule;
[0020] The classification output submodule accurately classifies the sample according to the received category data value and outputs the final detection and recognition result.
[0021] As a preferred solution for a SAR scene target recognition method based on a lightweight neural network, in the process of constructing the training data set, the training data set is the SAR scene target signal obtained by inverse imaging after fusing the MSTAR data set with the MiniSAR data set through a CM image fusion strategy.
[0022] The present invention also provides a SAR scene target recognition device based on a lightweight neural network, based on the above SAR scene target recognition method based on a lightweight neural network, comprising:
[0023] A SAR scene target recognition system building module is used to build a SAR scene target recognition system based on a lightweight neural network, wherein the SAR scene target recognition system includes an input preprocessing submodule, a basic feature extraction submodule, a core feature extraction submodule, a deep feature extraction submodule, a spatial mapping submodule, and a classification output submodule;
[0024] A training data set construction module, used to construct a training data set, and train the SAR scene target recognition system through the training data set;
[0025] A SAR scene target recognition system training module, used to train the setting module of the SAR scene target recognition system through the training data set and save the weight parameters of the setting module after training;
[0026] A trained SAR scene target recognition system acquisition module, used to load the weight parameters trained by the setting module into the SAR scene target recognition system to obtain the trained SAR scene target recognition system;
[0027] The SAR scene target recognition system recognition application module is used to input the SAR scene target signal into the trained SAR scene target recognition system, perform recognition through the trained SAR scene target recognition system, and obtain a final recognition result.
[0028] As a preferred solution of a SAR scene target recognition device based on a lightweight neural network, in the SAR scene target recognition system building module, the input preprocessing submodule is used to annotate the training data set and divide the training data set into a training set and a test set; by randomly scattering the samples, the amplitude, phase and wavelet domain information of the samples are extracted;
[0029] The basic feature extraction submodule extracts edge information of SAR scene targets and understands the shape and contour of the targets through the edge information.
[0030] As a preferred solution of a SAR scene target recognition device based on a lightweight neural network, in the SAR scene target recognition system building module, the core feature extraction submodule is used to extract local connection information of the SAR scene target, and reveal the local structure and relationship inside the target through the local connection information;
[0031] The deep feature extraction submodule is used to extract deep internal structure information, and capture high-level properties of SAR scene targets through the internal structure information.
[0032] As a preferred solution of a SAR scene target recognition device based on a lightweight neural network, in the SAR scene target recognition system building module, the spatial mapping submodule uses a multi-layer perceptron network and a spatial mapping strategy to perform weighted mapping on the features received from the deep feature extraction submodule to obtain a category data value, and sends the category data value to the classification output submodule;
[0033] The classification output submodule accurately classifies the sample according to the received category data value and outputs the final detection and recognition result.
[0034] As a preferred solution for a SAR scene target recognition device based on a lightweight neural network, in the training data set construction module, in the process of constructing the training data set, the training data set is a SAR scene target signal that is inversely imaged after fusing the MSTAR data set with the MiniSAR data set through a CM image fusion strategy.
[0035] The present invention has the following advantages: constructing a SAR scene target recognition system based on a lightweight neural network, the SAR scene target recognition system includes an input preprocessing submodule, a basic feature extraction submodule, a core feature extraction submodule, a deep feature extraction submodule, a spatial mapping submodule and a classification output submodule; constructing a training data set, and training the SAR scene target recognition system through the training data set; training the setting module of the SAR scene target recognition system through the training data set, and saving the weight parameters of the setting module after training; loading the weight parameters of the setting module after training into the SAR scene target recognition system to obtain the trained SAR scene target recognition system; inputting the SAR scene target signal into the trained SAR scene target recognition system, and identifying through the trained SAR scene target recognition system to obtain the final recognition result. The present invention is based on the MFW-Light (Multiple-Input Multiple-Output Fusion Wavelet Transform Neural Networks-Lightweight) previously studied by the authors, and through carefully adjusting the network structure, the original 8 residual blocks are reduced to 6, and the network parameters are reduced, thereby obtaining a multi-input lightweight neural network. Under the premise of maintaining the same hyperparameter settings, the multi-input lightweight neural network not only ensures a recognition accuracy of more than 93%, but also achieves a 1.1-fold increase in training speed, which makes it more practical in emergency scenarios in actual engineering. At the same time, the SAR scene target recognition method used by the multi-input lightweight neural network takes the amplitude, phase and wavelet domain information of the SAR signal as input. This strategy enables us to deeply explore and fully utilize the diverse features of SAR images, thereby significantly improving the recognition performance and injecting new vitality into the field of SAR image recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.
[0037] The structures, proportions, sizes, etc. illustrated in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with the technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantial technical significance. Any structural modification, change in proportion or adjustment of size shall still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.
[0038] Figure 1 This is a flow chart of a SAR scene target recognition method based on a lightweight neural network provided in Example 1 of the present invention;
[0039] Figure 2 This is a logical structure diagram of a SAR scene target recognition method based on a lightweight neural network provided in Example 1 of the present invention;
[0040] Figure 3 Partial images of a data set used in a SAR scene target recognition method based on a lightweight neural network provided in Example 1 of the present invention (only the processed SAR grayscale image is shown);
[0041] Figure 4 A local magnified image of a portion of an image in a data set used in a SAR scene target recognition method based on a lightweight neural network provided in Example 1 of the present invention (only the processed SAR grayscale image is shown);
[0042] Figure 5 This is a schematic diagram of typical test results of a multi-input lightweight neural network proposed in a SAR scene target recognition method based on a lightweight neural network provided in Example 1 of the present invention;
[0043] Figure 6 This is a schematic diagram of typical test results of MFW-Light in a SAR scene target recognition method based on a lightweight neural network provided in Example 1 of the present invention;
[0044] Figure 7 This is a schematic diagram of the architecture of a SAR scene target recognition device based on a lightweight neural network provided in Example 2 of the present invention. DETAILED DESCRIPTION
[0045] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0046] Example 1
[0047] See also Figure 1 Embodiment 1 of the present invention provides a SAR scene target recognition method based on a lightweight neural network, comprising the following steps:
[0048] S1. Constructing a SAR scene target recognition system based on a lightweight neural network, wherein the SAR scene target recognition system includes an input preprocessing submodule, a basic feature extraction submodule, a core feature extraction submodule, a deep feature extraction submodule, a spatial mapping submodule and a classification output submodule;
[0049] S2. Constructing a training data set, and training the SAR scene target recognition system by using the training data set;
[0050] S3, training the setting module of the SAR scene target recognition system through the training data set, and saving the weight parameters of the setting module after training;
[0051] S4, loading the weight parameters trained by the setting module into the SAR scene target recognition system to obtain the trained SAR scene target recognition system;
[0052] S5. Input the SAR scene target signal into the trained SAR scene target recognition system, and perform recognition through the trained SAR scene target recognition system to obtain a final recognition result.
[0053] See also Figure 2 In this embodiment, in step S1, a SAR scene target recognition system based on a lightweight neural network is constructed, and the SAR scene target recognition system includes an input preprocessing submodule, a basic feature extraction submodule, a core feature extraction submodule, a deep feature extraction submodule, a spatial mapping submodule and a classification output submodule;
[0054] Specifically, a SAR scene target recognition system based on a multi-input lightweight neural network is constructed, which includes an input preprocessing submodule, a basic feature extraction submodule, a core feature extraction submodule, a deep feature extraction submodule, a spatial mapping submodule and a classification output submodule; the SAR scene target recognition system retains the essential characteristics of the SAR signal by taking the amplitude and phase of the SAR signal and the wavelet domain information as input, and at the same time combines the multi-input lightweight neural network to facilitate the extraction of the characteristics of the SAR scene target, thereby better improving the recognition performance of the multi-input lightweight neural network and improving the SAR scene target recognition accuracy of the SAR.
[0055] The input preprocessing submodule is used to annotate the training data set and divide the training data set into a training set and a test set; extract the amplitude, phase and wavelet domain information of the samples by randomly scattering the samples;
[0056] Specifically, the input preprocessing submodule is responsible for labeling the data set, dividing it into a training set and a test set, and randomly breaking up the classified samples to enhance the generalization ability of the network. In addition, the input preprocessing submodule also applies random additive and multiplicative noise with a signal-to-noise ratio of 25dB to the samples of the input system. This data enhancement strategy not only improves the utilization rate of the samples, but also enhances the robustness of the model to noise. Then, the amplitude, phase and wavelet domain information of the samples are extracted through the amplitude input processing branch, the phase input processing branch, and the wavelet domain input processing branch. At the same time, the input data is standardized, which helps to speed up the system to find the optimal parameters, thereby improving the overall training efficiency and model performance.
[0057] The basic feature extraction submodule extracts edge information of SAR scene targets and understands the shape and contour of the targets through the edge information.
[0058] Specifically, the basic feature extraction submodule is dedicated to extracting key edge information features from the input SAR scene target. In order to overcome the gradient vanishing problem that may occur in deep network training, the module design adopts residual blocks, which are based on the concept of identity mapping and effectively improve the efficiency of feature extraction and the overall performance of the network.
[0059] The core feature extraction submodule is used to extract local connection information of SAR scene targets, and to reveal the local structure and relationship inside the target through the local connection information;
[0060] Specifically, the core feature extraction submodule focuses on extracting local contact information features of input SAR scene targets. By increasing the number of residual blocks, the core feature extraction submodule's ability to capture feature details and overall feature extraction efficiency are significantly improved.
[0061] The deep feature extraction submodule is used to extract deep internal structure information, and capture high-level properties of SAR scene targets through the internal structure information.
[0062] Specifically, the deep feature extraction submodule focuses on extracting internal structural information features from the input SAR scene target. By integrating a variety of different residual blocks, the deep feature extraction submodule enhances its ability to capture complex features, thereby significantly improving the depth and efficiency of feature extraction.
[0063] The spatial mapping submodule uses a spatial mapping strategy to perform weighted mapping on the features received from the deep feature extraction submodule through a multi-layer perceptron network to obtain a category data value, and sends the category data value to the classification output submodule;
[0064] Specifically, the spatial mapping submodule uses a multi-layer perceptron network to adopt a spatial mapping method to perform weighted mapping on the features received from the deep feature extraction module to obtain category data values.
[0065] The classification output submodule accurately classifies the samples according to the received category data values and outputs the final detection and recognition results.
[0066] Specifically, the classification output submodule uses the category data value to accurately classify the sample and output the final detection and recognition result.
[0067] In this embodiment, in step S2, in the process of constructing the training data set, the training data set is the SAR scene target signal obtained by inverse imaging after fusing the MSTAR data set with the MiniSAR data set through the CM image fusion strategy.
[0068] Specifically, the specific steps of the CM image fusion strategy are:
[0069] T1, select the N×N matrix A in the MSTAR dataset image;
[0070] T2, select the M×M (M>N) matrix B in the MiniSAR dataset image;
[0071] T3. Multiply the matrix A and the CM matrix (matrix C) to obtain the matrix D.
[0072] T4. Add the matrix D and the N×N matrix E in the matrix B to obtain the fused matrix F;
[0073] T5. According to steps T1-T4, all images in the MSTAR dataset are fused with one scene in the MiniSAR dataset, and then a new dataset is obtained through inverse imaging.
[0074] The CM matrix is:
[0075]
[0076] Taking the outermost circle of the matrix as the first layer, the N×N matrix has Indicates rounding up. CM matrix The layers have the same value, and the number of i layers is determined according to the following formula.
[0077]
[0078] Where CM(i) represents the value of the CM matrix layer i. The value of the CM matrix layer i is similar to the tanh function, which makes the fused image transition naturally at the edge of the original two images.
[0079] Some images of the new dataset are as follows Figure 3 As shown, the local enlarged picture is as follows Figure 4 shown.
[0080] In this embodiment, in step S3, the setting module of the SAR scene target recognition system is trained by the training data set, and the weight parameters of the setting module after training are saved;
[0081] Specifically, the training data set is used to train the basic feature extraction submodule, core feature extraction submodule, deep feature extraction submodule and spatial mapping submodule of the SAR scene target recognition system, and the weight parameters of the trained basic feature extraction submodule, core feature extraction submodule, deep feature extraction submodule and spatial mapping submodule are saved.
[0082] The specific steps are as follows:
[0083] A1. Initialize weight parameters. Use a normal distribution with a mean of 0 and a variance of 0.01 to initialize the network weight parameters of the basic feature extraction submodule, the core feature extraction submodule, the deep feature extraction submodule, and the spatial mapping submodule.
[0084] A2. Set the training parameters of the SAR scene target recognition system. Set the batch size (batch_size) to 256 and the maximum training step (maxepoch) to 20; use AdamW as the model training optimizer, set the hyperparameters β1 to 0.95, β2 to 0.9, and weight decay to 0.001;
[0085] A3. Initialize the training iteration parameter epoch=1, initialize the batch iteration parameter num_batch=0 (indicates that the current batch is the num_batch batch in the current epoch), and initialize the error value Loss of the first round of training=0;
[0086] A4, the input preprocessing submodule receives a batch of SAR scene target signal data from the training set, applies random additive and multiplicative noise with a signal-to-noise ratio of 25 dB, extracts the amplitude, phase and wavelet domain information of the SAR scene target signal data, obtains the preprocessed SAR scene target data, and sends the preprocessed SAR scene target data to the basic feature extraction module;
[0087] A5, the basic feature extraction submodule extracts the key edge information features of the pre-processed SAR scene target data, and then inputs them into the core feature extraction submodule;
[0088] A6, the core feature extraction submodule extracts the local contact information features of the SAR scene target data, and then inputs them to the deep feature extraction submodule;
[0089] A7, the deep feature extraction submodule extracts the internal structure information features of the SAR scene target data, and then inputs them into the spatial mapping submodule;
[0090] A8, the spatial mapping submodule uses a multi-layer perceptron network to perform weighted processing on the features received from the deep feature extraction submodule to obtain category data values;
[0091] A9, the classification output submodule uses the softmax activation function to activate the category data value received from the spatial mapping submodule to obtain the predicted value of the category to which the sample belongs;
[0092] A10, the classification output submodule calculates the loss function value Loss between the predicted value and the true value through the cross entropy function. The calculation formula is:
[0093]
[0094] In the formula, p i represents the probability that the input SAR scene target data belongs to the i-th category of SAR scene target type; y i is the sample label; I is the total number of SAR scene target data types;
[0095] A11. Use gradient back propagation to update the network weight parameters of the SAR scene target recognition system based on multi-input lightweight neural network;
[0096] A12. When num_batch is equal to 15, epoch increases by 1;
[0097] A13, determine whether the training is finished. If the Loss calculated in step A10 is less than or equal to 0.05, the training is finished. If epoch is less than maxepoch and the Loss calculated in step A10 is greater than 0.05, go to step A4 to continue training. If epoch is equal to maxepoch, it means that the training is finished, and the weight parameters of the basic feature extraction module, the core feature extraction module and the deep feature extraction module after the current epoch are saved.
[0098] In this embodiment, in step S4, the SAR scene target recognition system loads pre-stored weight parameters to obtain the trained SAR scene target recognition system.
[0099] In this embodiment, in step S5, the SAR scene target signal is input into the trained SAR scene target recognition system, and the trained SAR scene target recognition system performs recognition to obtain a final recognition result.
[0100] Specifically, the trained SAR scene target recognition system uses an input preprocessing submodule, a basic feature extraction submodule, a core feature extraction submodule, a deep feature extraction submodule, a spatial mapping submodule and a classification output submodule to detect and identify the type of the input SAR scene target signal to obtain the final result.
[0101] After rigorous testing, the SAR scene target recognition system achieved a classification accuracy of 93% on both the cross-validation set and the test set, and demonstrated excellent convergence stability. This result verifies the excellent performance of the present invention in SAR scene target type recognition.
[0102] The present invention has verified its good SAR scene target recognition effect through simulation test. Its typical training curve and final test accuracy are shown in Figure 5 As shown in Figure 1, the classification accuracy of SAR scene target recognition by the multi-input lightweight neural network can reach 93% on the cross-validation set and 93% on the additional test set. The MFW-Light training curve and the final test accuracy are shown in Figure 1. Figure 6 As shown in Table 1, the classification accuracy of MFW-Light for SAR scene target recognition on the cross-validation set is 98%, and the classification accuracy on the additional test set is also 98%. As shown in Table 1, the training time of MFW-Light is 36 minutes and 23 seconds, while the training time of the multi-input lightweight neural network is 16 minutes and 58 seconds.
[0103]
[0104] Table 1 Training duration
[0105] Therefore, in practical applications, the multi-input lightweight neural network has shown significant advantages over MFW-Light. While maintaining the same hyperparameter configuration, the multi-input lightweight neural network not only ensures a high recognition accuracy of more than 93%, but also further increases the training speed by 1.1 times. This achievement fully verifies the practicality and efficiency of the present invention.
[0106] In summary, the present invention constructs a SAR scene target recognition system based on a lightweight neural network, and the SAR scene target recognition system includes an input preprocessing submodule, a basic feature extraction submodule, a core feature extraction submodule, a deep feature extraction submodule, a spatial mapping submodule and a classification output submodule; constructs a training data set, and trains the SAR scene target recognition system through the training data set; trains the setting module of the SAR scene target recognition system through the training data set, and saves the weight parameters of the setting module after training; loads the weight parameters of the setting module after training into the SAR scene target recognition system to obtain the trained SAR scene target recognition system; inputs the SAR scene target signal into the trained SAR scene target recognition system, and recognizes it through the trained SAR scene target recognition system to obtain the final recognition result. The present invention is based on the MFW-Light (Multiple-Input Multiple-Output Fusion Wavelet Transform Neural Networks-Lightweight) previously studied by the authors. By carefully adjusting the network structure, the original 8 residual blocks are reduced to 6, and the network parameters are reduced, thereby obtaining a multi-input lightweight neural network. Under the premise of maintaining the same hyperparameter settings, the multi-input lightweight neural network not only ensures a recognition accuracy of more than 93%, but also achieves a 1.1-fold increase in training speed, which makes it more practical in emergency scenarios in actual engineering. At the same time, the SAR scene target recognition method used by the multi-input lightweight neural network takes the amplitude, phase and wavelet domain information of the SAR signal as input. This strategy enables us to deeply explore and fully utilize the diverse features of SAR images, thereby significantly improving the recognition performance and injecting new vitality into the field of SAR image recognition.
[0107] It should be noted that the method of the embodiment of the present disclosure can be performed by a single device, such as a computer or a server. The method of the present embodiment can also be applied in a distributed scenario and completed by multiple devices cooperating with each other. In the case of such a distributed scenario, one of the multiple devices can only perform one or more steps in the method of the embodiment of the present disclosure, and the multiple devices will interact with each other to complete the described method.
[0108] It should be noted that the above describes some embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0109] Example 2
[0110] See also Figure 7 Embodiment 2 of the present invention further provides a SAR scene target recognition device based on a lightweight neural network, comprising:
[0111] SAR scene target recognition system construction module 001, used to construct a SAR scene target recognition system based on a lightweight neural network, the SAR scene target recognition system includes an input preprocessing submodule, a basic feature extraction submodule, a core feature extraction submodule, a deep feature extraction submodule, a spatial mapping submodule and a classification output submodule;
[0112] A training data set construction module 002 is used to construct a training data set, and train the SAR scene target recognition system through the training data set;
[0113] A SAR scene target recognition system training module 003 is used to train the setting module of the SAR scene target recognition system through the training data set and save the weight parameters of the setting module after training;
[0114] A trained SAR scene target recognition system acquisition module 004 is used to load the weight parameters trained by the setting module into the SAR scene target recognition system to obtain the trained SAR scene target recognition system;
[0115] The SAR scene target recognition system recognition application module 005 is used to input the SAR scene target signal into the trained SAR scene target recognition system, perform recognition through the trained SAR scene target recognition system, and obtain a final recognition result.
[0116] In this embodiment, in the SAR scene target recognition system construction module 001, the input preprocessing submodule is used to annotate the training data set and divide the training data set into a training set and a test set; by randomly scattering the samples, the amplitude, phase and wavelet domain information of the samples are extracted;
[0117] The basic feature extraction submodule extracts edge information of SAR scene targets and understands the shape and contour of the targets through the edge information.
[0118] In this embodiment, in the SAR scene target recognition system construction module 001, the core feature extraction submodule is used to extract local connection information of the SAR scene target, and reveal the local structure and relationship inside the target through the local connection information;
[0119] The deep feature extraction submodule is used to extract deep internal structure information, and capture high-level properties of SAR scene targets through the internal structure information.
[0120] In this embodiment, in the SAR scene target recognition system building module 001, the spatial mapping submodule uses a multi-layer perceptron network and a spatial mapping strategy to perform weighted mapping on the features received from the deep feature extraction submodule to obtain a category data value, and sends the category data value to the classification output submodule;
[0121] The classification output submodule accurately classifies the samples according to the received category data values and outputs the final detection and recognition results.
[0122] In this embodiment, in the training data set construction module 002, in the process of constructing the training data set, the training data set is the SAR scene target signal obtained by inverse imaging after fusing the MSTAR data set with the MiniSAR data set through the CM image fusion strategy.
[0123] It should be noted that the information interaction, execution process and other contents between the modules of the above-mentioned system are based on the same concept as the method embodiment in Example 1 of the present application, and the technical effects they bring are the same as those of the method embodiment of the present application. For specific contents, please refer to the description in the method embodiment shown above in the present application, and will not be repeated here.
[0124] Example 3
[0125] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium, in which a program code of a SAR scene target recognition method based on a lightweight neural network is stored, and the program code includes instructions for executing the SAR scene target recognition method based on a lightweight neural network of embodiment 1 or any possible implementation thereof.
[0126] The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that includes one or more available media. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).
[0127] Example 4
[0128] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;
[0129] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute a SAR scene target recognition method based on a lightweight neural network in Example 1 or any possible implementation thereof.
[0130] Specifically, the processor can be implemented by hardware or by software. When implemented by hardware, the processor can be a logic circuit, an integrated circuit, etc.; when implemented by software, the processor can be a general-purpose processor implemented by reading software codes stored in a memory. The memory can be integrated in the processor or can be located outside the processor and exist independently.
[0131] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable systems. The computer instructions can be stored in a computer-readable storage medium, or transmitted from a computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center.
[0132] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing system, they can be concentrated on a single computing system, or distributed on a network composed of multiple computing systems, and optionally, they can be implemented by a program code executable by a computing system, so that they can be stored in a storage system and executed by the computing system, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.
[0133] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.
Claims
1. A SAR scene target recognition method based on lightweight neural network, characterized in that: include: Constructing a SAR scene target recognition system based on a lightweight neural network, the SAR scene target recognition system includes an input preprocessing submodule, a basic feature extraction submodule, a core feature extraction submodule, a deep feature extraction submodule, a spatial mapping submodule and a classification output submodule; Constructing a training data set, and training the SAR scene target recognition system through the training data set; Training a setting module of the SAR scene target recognition system through the training data set, and saving weight parameters of the setting module after training; Loading the weight parameters trained by the setting module into the SAR scene target recognition system to obtain the trained SAR scene target recognition system; The SAR scene target signal is input into the trained SAR scene target recognition system, and the trained SAR scene target recognition system is used to perform recognition to obtain a final recognition result.
2. The SAR scene target recognition method based on lightweight neural network according to claim 1 is characterized in that: The input preprocessing submodule is used to annotate the training data set and divide the training data set into a training set and a test set; extract the amplitude, phase and wavelet domain information of the samples by randomly scattering the samples; The basic feature extraction submodule extracts edge information of SAR scene targets and understands the shape and contour of the targets through the edge information.
3. The SAR scene target recognition method based on lightweight neural network according to claim 2 is characterized in that: The core feature extraction submodule is used to extract local connection information of SAR scene targets, and to reveal the local structure and relationship inside the target through the local connection information; The deep feature extraction submodule is used to extract deep internal structure information, and capture high-level properties of SAR scene targets through the internal structure information.
4. The SAR scene target recognition method based on lightweight neural network according to claim 3 is characterized in that: The spatial mapping submodule uses a spatial mapping strategy to perform weighted mapping on the features received from the deep feature extraction submodule through a multi-layer perceptron network to obtain a category data value, and sends the category data value to the classification output submodule; The classification output submodule accurately classifies the sample according to the received category data value and outputs the final detection and recognition result.
5. The SAR scene target recognition method based on lightweight neural network according to claim 4 is characterized in that: In the process of constructing the training data set, the training data set is the SAR scene target signal obtained by inverse imaging after fusing the MSTAR data set with the MiniSAR data set through a CM image fusion strategy.
6. A SAR scene target recognition device based on a lightweight neural network, using any one of the SAR scene target recognition methods based on a lightweight neural network according to claims 1-5, characterized in that: include: A SAR scene target recognition system building module is used to build a SAR scene target recognition system based on a lightweight neural network, wherein the SAR scene target recognition system includes an input preprocessing submodule, a basic feature extraction submodule, a core feature extraction submodule, a deep feature extraction submodule, a spatial mapping submodule, and a classification output submodule; A training data set construction module, used to construct a training data set, and train the SAR scene target recognition system through the training data set; A SAR scene target recognition system training module, used to train the setting module of the SAR scene target recognition system through the training data set and save the weight parameters of the setting module after training; A trained SAR scene target recognition system acquisition module, used to load the weight parameters trained by the setting module into the SAR scene target recognition system to obtain the trained SAR scene target recognition system; The SAR scene target recognition system recognition application module is used to input the SAR scene target signal into the trained SAR scene target recognition system, perform recognition through the trained SAR scene target recognition system, and obtain a final recognition result.
7. The SAR scene target recognition device based on lightweight neural network according to claim 6, characterized in that: In the SAR scene target recognition system building module, the input preprocessing submodule is used to annotate the training data set and divide the training data set into a training set and a test set; extract the amplitude, phase and wavelet domain information of the samples by randomly scattering the samples; The basic feature extraction submodule extracts edge information of SAR scene targets and understands the shape and contour of the targets through the edge information.
8. The SAR scene target recognition device based on lightweight neural network according to claim 7, characterized in that: In the SAR scene target recognition system building module, the core feature extraction submodule is used to extract local connection information of the SAR scene target, and reveal the local structure and relationship inside the target through the local connection information; The deep feature extraction submodule is used to extract deep internal structure information, and capture high-level properties of SAR scene targets through the internal structure information.
9. The SAR scene target recognition device based on lightweight neural network according to claim 8, characterized in that: In the SAR scene target recognition system building module, the spatial mapping submodule uses a multi-layer perceptron network and a spatial mapping strategy to perform weighted mapping on the features received from the deep feature extraction submodule to obtain a category data value, and sends the category data value to the classification output submodule; The classification output submodule accurately classifies the sample according to the received category data value and outputs the final detection and recognition result.
10. The SAR scene target recognition device based on lightweight neural network according to claim 9, characterized in that: In the training data set construction module, in the process of constructing the training data set, the training data set is the SAR scene target signal obtained by inverse imaging after fusing the MSTAR data set with the MiniSAR data set through the CM image fusion strategy.
Citation Information
Cited By
SAR (Synthetic Aperture Radar) scene target generation method and device based on spatial calibration
CN119575336A
