A joint SAR target recognition method in data domain and image domain based on ResNet

Abstract

This invention discloses a joint SAR target recognition method based on ResNet in the data and image domains, relating to the field of radar signal processing technology. It proposes a joint target recognition method based on synthetic aperture radar echo data and image data. The common approach to imaging radar target recognition technology is to obtain radar images of the target using imaging algorithms, and then perform image-based target recognition. However, imaging radar echoes contain complete target information, while imaging algorithms may lose target information due to performance limitations. Therefore, this invention, based on a residual neural network, independently performs recognition on echo data and image data as inputs, obtaining recognition results. Then, it utilizes Dempster-Shafer evidence theory for decision-level fusion to make a target category decision, effectively leveraging the advantages of both echo-based and image-based recognition technologies, thereby significantly improving the overall target recognition accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of radar signal processing, and in particular relates to a data domain and image domain joint SAR target recognition method based on ResNet. Background Art

[0002] With the rapid development of modern information technology, target recognition technology has become an important technical support for radar intelligence and informatization, and has broad application prospects in the fields of early warning detection, precision guidance, etc. Synthetic Aperture Radar (SAR) and Inverse Synthetic Aperture Radar (ISAR) can obtain high-resolution images of stationary and moving targets, and have become very effective tools for identifying various targets. In recent years, with the development of big data technology and high-performance computing resources, intelligent target recognition uses artificial intelligence technology for feature extraction and sample accumulation to achieve cognition and identification of unknown targets. Different from other automatic target recognition methods, intelligent target recognition has the characteristics of autonomous learning and intelligent reasoning. Among them, intelligent target recognition with deep learning as the core has become an effective way to solve the problem of automatic target recognition. A large number of researchers have also studied some intelligent radar target recognition methods based on deep learning for SAR / ISAR images.

[0003] However, existing SAR / ISAR target recognition methods are basically based on two-dimensional images, that is, the working mode of imaging first and then recognition is adopted. In fact, the imaging radar echo contains complete target information, while the imaging algorithm may lose target information due to the limitation of algorithm performance. Summary of the invention

[0004] In order to solve the technical problems mentioned in the above background technology, the present invention proposes a data domain and image domain joint SAR target recognition method based on ResNet.

[0005] In order to achieve the above technical objectives, the technical solution of the present invention is:

[0006] A data domain and image domain joint SAR target recognition method based on ResNet, comprising the following steps:

[0007] S1. Construct the overall framework of the joint SAR target recognition model;

[0008] S2, establishing a target recognition module based on the overall framework constructed in step S1;

[0009] S3, independently identifying the echo data as input and the image data as input; for each target to be identified, two different target category memberships are obtained, thereby obtaining two basic probability distribution functions;

[0010] S4, fusing the basic probability assignments according to the DS combination rule to obtain the fused basic assignment probability;

[0011] S5. The trust function is obtained by using the basic assignment probability after fusion in step S4, and the category with the largest trust function is used as the established judgment logic to obtain the best matching category with the target to be identified.

[0012] Preferably, the overall framework of the joint SAR target recognition model in step S1 includes echo data, image data, a first residual neural network, a second residual neural network, basic probability assignment rules, evidence combination rules and discriminant logic rules, wherein the basic probability assignment rule obtains the basic probability distribution function through the category membership of each target to be identified; the evidence combination rule fuses the basic probability assignment through the DS combination rule; and the discriminant logic rule takes the category with the largest trust function as the fusion recognition result.

[0013] Preferably, in step S2, the architecture of the residual neural network includes first to third channels, convolution kernels, Relu layers, pooling layers, first to fourth convolution layers, fully connected layers, Dropout layers and Softmax layers, the first to third channels are respectively connected to the convolution kernels, the convolution kernels, Relu layers, pooling layers, first to fourth convolution layers, fully connected layers, Dropout layers and Softmax layers are connected in sequence, wherein the structures of the first to fourth convolution layers are the same, the first convolution layer includes a first convolution kernel, a first Relu unit, a second convolution kernel, a Shortcut connection unit, a summation unit, a second Relu unit and a pooling unit, the first convolution kernel, the first Relu unit, the second convolution kernel, the summation unit, the second Relu unit and the pooling unit are connected in sequence, the input end of the Shortcut connection unit is connected to the input end of the first convolution kernel, and the output end of the Shortcut connection unit is connected to the input end of the summation unit.

[0014] Preferably, in step S1, the basic probability assignment function m(.) is a function from 2 Ω →[0, 1], and satisfies and m(φ)=0, then m(.) is called the basic probability assignment function on the recognition framework Ω. At this time, the confidence function Bel formula based on the basic probability assignment function on the recognition framework Ω is expressed as follows:

[0015]

[0016] The likelihood function P1 based on the basic probability assignment function in the recognition framework Ω is expressed as follows:

[0017]

[0018] The confidence interval [Bel, P1] is used to represent the degree of confirmation of the subset A in the recognition framework Ω.

[0019] Preferably, step S3 refers to:

[0020] Let Ω be a recognition framework, denoted as Ω A ={A1, A2, ..., A N} and Ω B = {B1, B2, ..., B N}, where N is the number of categories. For each target to be identified, two different target category memberships are obtained, expressed as P image (A) = {P image (A1), P image (A2), ..., P image (A N )} and P echo (B) = {P echo (B1), P echo (B2), ..., P echo (B N )}, which contain information about N categories. The basic probability assignment function is denoted as m image (A) and m echo (B), taking image recognition as an example, its basic probability assignment m image (A) is obtained by the following formula:

[0021]

[0022]

[0023] Among them, μ is the uncertainty limit factor, which represents the confidence in the identification method used obtained by prior knowledge.

[0024] Preferably, in step S4, the basic probability assignment m obtained based on DS fusion is DS (F) The formula is expressed as follows:

[0025]

[0026] In the formula, 1-K is the normalization factor.

[0027] Preferably, the trust function Bel(A) and the decision rule formula in step S5 are expressed as follows:

[0028]

[0029]

[0030] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0031] (1) The present invention utilizes the ResNet residual neural network architecture with SAR echo data as input, breaking through the bottleneck of existing radar image recognition technology limited by imaging quality and improving recognition efficiency; compared with the previous convolutional neural network, the residual neural network has the advantages of deeper layers but lower complexity, and the network training is easier to optimize, and the recognition and classification performance is better.

[0032] (2) In addition, the method of the present invention utilizes the idea of ​​decision-level fusion. The overall architecture of target recognition is based on the DS evidence theory, which breaks through the target recognition technology of single modality data and effectively integrates the echo recognition results and image recognition results, thereby effectively improving the overall target recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the overall framework of the joint SAR target recognition model based on evidence theory;

[0034] Figure 2a This is a framework diagram of SAR target recognition based on ResNet. Figure 2b It is the expanded structure of the first to fourth convolutional layers;

[0035] Figure 3 The confusion matrix of the simulated ISAR data set obtained by the joint recognition method;

[0036] Figure 4 is the confusion matrix of the measured SAR data set obtained by the joint recognition method. DETAILED DESCRIPTION

[0037] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0038] The present invention proposes a data domain and image domain joint SAR target recognition method based on ResNet, comprising the following steps:

[0039] Step S1, constructing an overall framework of a joint SAR target recognition model based on evidence theory;

[0040] Step S2, constructing a SAR target recognition part based on ResNet, and independently performing recognition on echo data and image data as input;

[0041] Step S3: for each target to be identified, two different target category memberships are obtained respectively, thereby obtaining two basic probability distribution functions;

[0042] Step S4, fusing the basic probability assignment according to the DS combination rule to achieve decision-level fusion of echo data and image data, and obtaining the fused basic assignment probability;

[0043] Step S5: Obtain a trust function using the fused basic assignment probability, and use the category with the largest trust function as the established judgment logic to obtain the best matching category with the target to be identified;

[0044] The framework of SAR target recognition based on ResNet is shown in Figure 2. Figure 2(a) shows an overview of the ResNet structure, which consists of an input layer, convolution kernel, nonlinear activation layer, pooling layer, first to fourth convolution layers, fully connected layer, Dropout layer, and final output layer. Figure 2(b) shows the expanded structure of the first to fourth convolution layers, including convolution kernels, nonlinear activation layers, pooling layers, and shortcut connections.

[0045] ResNet was originally designed for image classification tasks, which require 3-channel real input data. However, SAR echo data is complex. Generally speaking, using the amplitude, real part, and imaginary part of complex data as three-channel input often yields the best results. Therefore, the data is input into the network in the same way, using the amplitude, real part, and imaginary part of the echo as three-channel input.

[0046] In the ResNet framework, the network uses a 3×3 convolution kernel, increasing the depth and number of non-linear activation layers. The non-linear activation part is the rectified linear unit (ReLU), which helps the network learn several times faster than other non-linear functions such as sigmoid or tanh. The pooling layer utilizes the maximum pooling layer for downsampling / local accumulation of feature maps to reduce the spatial dimension. A Dropout layer with a probability of 0.5 is used to deal with overfitting problems, which is usually caused by a small training dataset and large network parameters. The output layer is a Softmax classifier that predicts the probability distribution of different classes.

[0047] On the basis of the ordinary flattened network, residual units are added through shortcut connections to transform the network into a deep residual learning structure.

[0048] As an uncertainty reasoning method based on set theory, Dempster-shafer (DS) evidence theory can effectively handle conflicting information and provide a natural and powerful tool for the expression and synthesis of uncertain information. It can often be applied to the decision-level fusion recognition task of the target. Based on the fusion of the DS method, the data and information from multiple sensors and information sources are first preprocessed, and then the basic probability distribution function, credibility function, and likelihood function of multiple data are calculated. Then, according to the DS synthesis rule, the basic probability distribution function, credibility function, and likelihood function under the joint action of all evidence are calculated. Finally, according to certain judgment rules, the hypothesis with the largest credibility and likelihood is selected as the fusion result.

[0049] In evidence theory, all objects in target recognition are called recognition frames Ω, and the categories in the recognition frame are finite mutually exclusive discrete values ​​A∈2 Ω , each set is called a focal element A. The core problem of evidence theory is to calculate the probability that an element belongs to a subset of Ω based on the given information when the frame to be identified is known. The core of evidence theory is the Basic Probability Assignment (BPA). BPA and its fusion rules are defined as follows:

[0050] The basic probability assignment function m(.) is a function from 2 Ω →[0, 1], and satisfies and m(φ) = 0, then m(.) is called the basic probability distribution function on the recognition framework Ω, also known as the quality function.

[0051] The belief function Bel based on BPA in the recognition framework Ω is defined as:

[0052]

[0053] The likelihood function Pl based on BPA in the recognition framework Ω is defined as:

[0054]

[0055] In evidence theory, the confidence interval [Bel, Pl] is used to represent the degree of confirmation of a subset A in the identification framework Ω.

[0056] The evidence theory approach has a reasonable information synthesis rule, called DS 2 Rules, usually used Symbolically, 2 Ω The DS rule for the two basic probability distribution functions m1(.) and m2(.) on is defined as: if m DS (φ)=0, for B,C∈2 Ω ,

[0057]

[0058] Among them, the coefficient 1 / (1-K) is the normalization factor, The value of K reflects the degree of conflict between the two pieces of evidence, usually referred to as the conflict coefficient, and its range is [0, 1]. If K≠1, then m(.) determines the basic probability distribution; if K=1, then m1 is considered to contradict m2, and the basic probability distribution cannot be merged. The closer K is to 1, the greater the conflict between the two pieces of evidence, and vice versa. Different pieces of evidence represent different sources of information. The basic probability distribution functions of the two systems represent the degree of support for each proposition by different systems, and the DS rule reflects the redistribution of information.

[0059] Based on the DS evidence theory, a joint SAR target recognition framework is constructed, such as Figure 1 As shown, the echo recognition result is fused with the image recognition result. The specific contents are as follows:

[0060] First, target recognition based on residual neural network is performed on image data and echo data respectively. Let Ω be a recognition framework, represented by Ω A ={A1, A2, ..., A N} and Ω B = {B1, B2, ..., B N}, where N is the number of categories. For each target to be identified, two different target category memberships are obtained, expressed as P image (A) = {P image (A1), P image (A2), ..., P image (A N )} and P echo (B) = {P echo (B1), P echo (B2), ..., P echo (B N )}, which contain information about N categories. The basic probability assignment function is denoted as m image (A) and m echo (B), taking image recognition as an example, its basic probability assignment m image (A) is obtained by the following formula:

[0061]

[0062]

[0063] Among them, μ is the uncertainty limit factor, which indicates the confidence in the identification method used based on prior knowledge. The higher the confidence, the smaller the uncertainty and μ. Generally speaking, μ≤1. The specific value needs to be appropriately selected according to the credibility of the evidence.

[0064] After obtaining the basic probability of evidence, the basic probability assignment is combined based on the DS combination rule, and the orthogonal sum of the basic probability of evidence and the normalization of the reasonable combination set are used to concentrate the trust on the elements with small uncertainty, thereby reducing the uncertainty of the judgment. DS (F) can be expressed as:

[0065]

[0066] In the formula, 1-K is the normalization factor.

[0067] Finally, according to the established judgment logic, the best matching category with the target to be identified is obtained. Generally, a relatively simple decision rule is adopted, and the category with the largest trust function is taken as the fusion recognition result. The trust function Bel(A) and the decision rule can be expressed as follows:

[0068]

[0069]

[0070] Implementation Examples

[0071] In order to verify the effectiveness of the ResNet-based joint SAR target recognition method in data domain and image domain, experimental verification was carried out on the ISAR simulation dataset and the SAR measured dataset respectively.

[0072] The performance of joint target recognition is evaluated on four types of simulated space satellite targets of ISAR, as shown in Table 1. Compared with image-based recognition, recognition based on echo data has a higher recognition rate, and the accuracy of joint recognition is further improved through DS evidence theory fusion. Figure 3 The confusion matrix of the simulated ISAR data set obtained by the joint recognition method.

[0073] Table 1 Experimental results of simulated ISAR dataset

[0074] method Recognition rate (%) Echo recognition based on ResNet 91.1 Image recognition based on ResNet 95.5 Joint Identification Based on DS Evidence Theory 99.9

[0075] In order to verify the performance of joint SAR target recognition on the measured dataset, experiments were conducted on ten categories of SAR ground targets on the MSTAR dataset. The evaluation of this dataset is shown in Table 2. Figure 4 The confusion matrix of the measured SAR dataset obtained by joint identification is shown.

[0076] Table 2 Experimental results of measured SAR dataset

[0077] method Recognition rate (%) Echo recognition based on ResNet 90.5 Image recognition based on ResNet 90.1 Joint Identification Based on DS Evidence Theory 97.5

[0078] As can be seen from Table 2, the recognition rate of the SAR target recognition method based on echo data using ResNet reaches more than 90%, and the recognition performance is better than that of the SAR target recognition method based on image. The joint SAR target recognition using DS evidence theory can achieve the highest accuracy of 97.5%.

[0079] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0080] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0081] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0082] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0083] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0084] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A data domain and image domain joint SAR target recognition method based on ResNet, characterized in that: The following steps are involved: S1. Construct the overall framework of the joint SAR target recognition model; S2, establishing a target recognition module based on the overall framework constructed in step S1; S3, independently identifying the echo data as input and the image data as input; for each target to be identified, two different target category memberships are obtained, thereby obtaining two basic probability distribution functions; S4, fusing the basic probability assignments according to the DS combination rule to obtain the fused basic assignment probability; S5, using the basic assignment probability after fusion in step S4 to obtain a trust function, taking the category with the largest trust function as the established judgment logic, and obtaining the best matching category with the target to be identified; Step S3 means: Set Ω A ={A1,A2,…,A N } and Ω B = {B1, B2, …, B N }, where Ω represents the recognition framework and N is the number of categories; for each target to be identified, two different target category memberships are obtained, expressed as P image (A) = {P image (A1),P image (A2),…,P image (A N )} and P echo (B)= {P echo (B1),P echo (B2),…,P echo (B N )}, which contain information about N categories; the basic probability distribution function is represented by m image (A) and m echo (B), in image recognition, its basic probability assignment m image (A) is obtained by the following formula: Among them, μ is the uncertainty limit factor, which represents the confidence in the identification method used obtained by prior knowledge; The basic probability assignment m obtained based on DS fusion in step S4 DS (F) The formula is expressed as follows: In the formula, 1-K is the normalization factor.

2. According to claim 1, a data domain and image domain joint SAR target recognition method based on ResNet is characterized in that: The overall framework of the joint SAR target recognition model in step S1 includes echo data, image data, a first residual neural network, a second residual neural network, basic probability assignment rules, evidence combination rules and discriminant logic rules, wherein the basic probability assignment rule obtains the basic probability distribution function through the category membership of each target to be identified; the evidence combination rule fuses the basic probability assignment through the DS combination rule; and the discriminant logic rule takes the category with the largest trust function as the fusion recognition result.

3. According to claim 1, a data domain and image domain joint SAR target recognition method based on ResNet is characterized in that: In step S2, the architecture of the residual neural network includes first to third channels, convolution kernels, Relu layers, pooling layers, first to fourth convolution layers, fully connected layers, Dropout layers and Softmax layers, the first to third channels are connected to the convolution kernels respectively, the convolution kernels, Relu layers, pooling layers, first to fourth convolution layers, fully connected layers, Dropout layers and Softmax layers are connected in sequence, wherein the structures of the first to fourth convolution layers are the same, the first convolution layer includes a first convolution kernel, a first Relu unit, a second convolution kernel, a Shortcut connection unit, a summation unit, a second Relu unit and a pooling unit, the first convolution kernel, the first Relu unit, the second convolution kernel, the summation unit, the second Relu unit and the pooling unit are connected in sequence, the input end of the Shortcut connection unit is connected to the input end of the first convolution kernel, and the output end of the Shortcut connection unit is connected to the input end of the summation unit.

4. According to claim 3, a data domain and image domain joint SAR target recognition method based on ResNet is characterized in that: In step S1, the basic probability assignment function m(.) is a function from 2 Ω →[0,1], when it satisfies When m(φ) = 0, m(.) is called the basic probability assignment function on the recognition framework Ω. At this time, the confidence function Bel formula based on the basic probability assignment function on the recognition framework Ω is expressed as follows: In the formula, m(.) represents the basic probability assignment function, A and B represent different subsets in the recognition framework Ω; The likelihood function Pl based on the basic probability assignment function in the recognition framework Ω is expressed as follows: In the formula, m(.) represents the basic probability assignment function, A and B represent different subsets in the recognition framework Ω; The confidence interval [Bel, Pl] is used to represent the degree of confirmation of the subset A in the identification framework Ω.

5. According to claim 1, a data domain and image domain joint SAR target recognition method based on ResNet is characterized in that: The trust function Bel(A) and the decision rule formula in step S5 are expressed as follows: In the formula, m DS (F) represents the basic probability assignment obtained by fusion in step S4, P j is the probability that the target belongs to the jth class, Bel(P j ) is the trust function that the target belongs to the jth class, P DS is the probability after the judgment.

Images