Electromagnetic target recognition and imaging method based on multi-layer perception neural network and metamaterials

By combining multilayer perceptual neural networks with information metamaterials, the problems of low resolution and poor robustness in traditional microwave imaging systems have been solved, achieving efficient and noise-resistant electromagnetic target recognition and image reconstruction, suitable for imaging tasks in various environments.

CN114898005BActive Publication Date: 2024-11-05SOUTHEAST UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202210497134.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-09
Publication Date
2024-11-05
Estimated Expiration
2042-05-09

AI Technical Summary

Technical Problem

Traditional microwave imaging systems are limited by the observation aperture, resulting in low imaging resolution. High frequencies lead to high costs. Synthetic aperture technology is not suitable for imaging fixed ground targets. Furthermore, electromagnetic imaging systems using passive metamaterials have poor robustness, and machine learning performance is limited in noisy environments.

Method used

By combining multilayer perceptual neural networks with information metamaterials, feature space is extracted through the construction of multilayer perceptual neural networks, arbitrary beamforming in the electromagnetic space domain is performed using information metamaterials, and holographic projection inverse reconstruction and low-complexity image reconstruction algorithms are combined to achieve highly reliable and high-speed imaging target recognition.

Benefits of technology

It achieves high-accuracy target recognition and image reconstruction, improves imaging rate, enhances anti-noise interference capability, has strong system adaptability, and simplifies antenna structure.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114898005B_ABST
    Figure CN114898005B_ABST
Patent Text Reader

Abstract

The present invention discloses an electromagnetic target recognition imaging method based on a multi-layer perception neural network and metamaterials. The electromagnetic imaging system is composed of an excitation source antenna, an information metamaterial, an observed target and a receiving antenna; the information metamaterial is composed of N×N single-bit programmable phase control units, which perform the primary reflection surface control function, and all units can be synchronously and independently controlled in real time through a programmable gate array. In the pre-imaging preparation stage, an MLP network is constructed for feature learning, and a coding combination is generated through a holographic projection reconstruction algorithm. In the electromagnetic imaging link, the information metamaterial periodically executes the aforementioned coding combination, and at each time slot, the receiving end samples the secondary reflection signal reflected by the observed target. After a set of periodic sampling is completed, the receiving end reconstructs the electromagnetic information of the observed target based on the collected data group. The intelligent electromagnetic imaging system adaptively samples the sample set, effectively improving the imaging speed and imaging resolution, and significantly improving the anti-noise ability of imaging.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to an electromagnetic target recognition imaging method, in particular to an electromagnetic target recognition imaging method based on a multi-layer perception neural network and information metamaterials. Background Art

[0002] Traditional optical imaging systems are limited by the electromagnetic wave characteristics of the optical frequency band, and show certain limitations in optical instrument manufacturing, line-of-sight interference, and all-weather imaging. As a microwave imaging and detection system that complements optical imaging, it has the advantages of high penetration, anti-camouflage color, and simple structure. It has been widely used in security inspection, satellite imaging and other fields. However, since the wavelength of the microwave band is several orders of magnitude higher than that of the optical band, the resolution of the imaging result is limited by the equivalent wavelength size of the observation aperture. A variety of methods have been proposed to overcome this defect. The first method is to increase the beam frequency. The current detection frequency can reach THz, but the higher frequency brings higher device costs. Phased array technology is also used for beamforming and scanning, but each channel of the same array requires an independent RF and data processing link, which lacks practicality in portable imaging systems. In addition, synthetic aperture technology (SAR) is also used in satellite space-based imaging, but this method requires physically moving feed and detection antennas, which is not suitable for imaging fixed targets on the ground.

[0003] Electromagnetic metamaterials have brought new ideas to the field of electromagnetic imaging. Electromagnetic metamaterials (Metamaterials) are artificial materials composed of macroscopic basic units with specific geometric shapes arranged periodically or non-periodically. With the help of the basic units to control the phase or amplitude of electromagnetic waves in sub-wavelength space, the entire array / structure shows macroscopic control. Based on this, people can design artificial surface structures to artificially control this amplitude or phase discontinuity, and then use two-dimensional metasurfaces to control spatial propagation waves. Achieve the purpose of arbitrarily controlling reflected waves and refracted waves. At present, electromagnetic metamaterials have been widely used in beamforming, high-performance lenses, and beam synthesis. Among them, the concept of information metamaterials was proposed by Cui Tiejun's team in 2014. By equating the state of active adjustable units to quantized digital information, important concepts such as information entropy in the field of digital processing are introduced into the electromagnetic metamaterial system, connecting the digital world with the physical world, and greatly expanding the theoretical basis and application field of metamaterials.

[0004] At the same time, machine learning has been booming in recent years. Feature recognition technology makes it possible to reconstruct image information or identify image targets with only a few samples. Dataset-driven training methods can autonomously mine abstract features of targets without prior knowledge or human guidance.

[0005] Traditional metamaterial-based electromagnetic imaging recognition systems use passive metamaterials as electromagnetic wave control units. In order to obtain high-dimensional measurement data information, frequency diversity methods are often used, that is, collecting the reflection information of the observed target to the electromagnetic wave in a wide band. However, due to the limitations of passive metamaterials, the equivalent beam in the electromagnetic space domain can be considered as a stray random signal. In this way, machine learning methods are limited to the decoding field. In particular, the algorithm performance is severely affected by noise and the system is not robust. Summary of the invention

[0006] Technical problem: To overcome this problem, the present invention provides an electromagnetic target recognition imaging method based on a multi-layer perception neural network and metamaterials, which uses an information metamaterial device and can realize arbitrary beamforming in the electromagnetic space domain by virtue of its programmable ability. The multi-layer perception neural network MLP, as a classic algorithm in machine learning neural networks, can autonomously extract feature information from data sets. Therefore, by integrating the feature compression technology of electromagnetic space domain images under information metamaterials, high-reliability and high-speed imaging target recognition or image reconstruction tasks can be achieved.

[0007] Technical solution: The present invention discloses an electromagnetic target recognition imaging method based on a multi-layer perception neural network and metamaterials, comprising the following steps:

[0008] Step 101: Based on the imaging target potential sample set data, a multi-layer perceptron neural network MLP is constructed and trained on the selected sample set, and then a pattern feature space is extracted from the trained multi-layer perceptron neural network;

[0009] Step 102: Based on the feature space, holographic projection inverse reconstruction is performed on the feature vectors or patterns one by one, so that the primary reflected wavefront reflected by the feed source after passing through the information metasurface has a characteristic pattern field intensity distribution, thereby calculating and generating a set of coding combinations of the information metasurface;

[0010] Step 103: Establish an imaging system consisting of a feed antenna, an information metamaterial, an observation object platform, a receiving antenna and other devices, and collect recognition imaging data of the observation target;

[0011] Step 104: establishing a numerical model describing the intrinsic relationship between the target electromagnetic information and the receiving end signal in the imaging system constructed in step 103 based on the electromagnetic field electromagnetic wave space propagation theory;

[0012] Step 105: Based on the intrinsic relationship between the received electromagnetic wave signal and the electromagnetic information of the observed target in the numerical model derived in step 104, a low-complexity linear algorithm for image reconstruction is designed.

[0013] in,

[0014] The imaging target potential sample set described in step 101 is extracted based on the feature space of the multi-layer perceptron neural network MLP; in the MLP network model, the relationship between the i-th layer neurons and the i+1-th layer neurons is:

[0015] a i+1 =F(W i ·a i ) i=1,2,…,M-1 (1)

[0016] Where M is the number of MLP network layers, a i , a i+1 are the neuron outputs of the i-th layer and the i+1-th layer, respectively, W i is the weight matrix between the i-th layer and the i+1-th layer, and F() is the nonlinear activation function;

[0017] The MLP network established for feature extraction of latent sample set data of imaging targets has the same number of input neurons as the number of pixels of the imaging target, and the number of output neurons is determined by the specific type of network. In any case, the number of neurons in each initial layer of the network should decrease layer by layer to achieve feature compression.

[0018] In the network trained by the optimization algorithm, the weight matrix W i All have been optimized to reflect the feature space of the sample set.

[0019] The number of output neurons is determined by the specific type of network. When the ultimate goal of the imaging system is to identify the target, the number of output neurons is the corresponding number of label categories; when the ultimate goal of the imaging system is to image the target, the number of output neurons should be the same as the number of pixels of the imaging target.

[0020] The weight matrix W i , the weight matrix W1 between the first layer and the second layer is the selected main feature space matrix. The main feature space matrix W1 can be expressed as

[0021] W1=[w1 w2 … w N ] T (2)

[0022] Where N is a positive integer, representing the number of rows of matrix W1, that is, the dimension of the feature space; T represents the transposition symbol; w1 to w N are N groups of feature column vectors; the length of the column vectors is K, where K is a positive integer representing the number of pixels of the imaging target.

[0023] In step 102, the holographic projection inverse reconstruction is performed on the feature vectors or patterns one by one;

[0024] Each set of feature vectors w constructed in step 101 iThe image matrix is ​​converted into the same size as the observed target. Based on its numerical intensity data, the information metasurface encoding is calculated using the holographic projection inverse reconstruction modified Gerchberg-Saxton (GS) algorithm, so that the primary reflected wavefront reflected by the feed source after passing through the information metasurface has a characteristic vector w i The distribution of the converted feature pattern field and the N groups of codes obtained by calculation are stored in the FPGA for subsequent calls.

[0025] The step 103 of establishing an imaging system composed of a feed antenna, an information metamaterial, an observation object platform, a receiving antenna and other devices includes the following specific steps:

[0026] Based on the position of the observation target (303), the metamaterial device (302) is placed in front of the observation target,

[0027] Based on the relative position between the observation target position and the metamaterial device, the excitation source antenna (301) is installed to ensure that the excitation source center and the observation target are mirror-symmetrical about the metamaterial device, and the main radiation direction of the excitation source antenna is aligned with the center of the metamaterial device, and the beam can cover the metamaterial device;

[0028] Based on the size and position of the observation target, a receiving antenna (305) is installed to ensure that the main radiation direction of the receiving antenna is aligned with the observation target and the receiving antenna is located in a relatively far field area of ​​the observation target;

[0029] Connecting the excitation source antenna to the radio frequency signal source generator, and connecting the receiving antenna to the radio frequency power meter (306); connecting the metamaterial device to the FPGA coding control device (304); connecting the host computer (307) to the radio frequency power meter and the coding control FPGA device respectively;

[0030] In this way, an intelligent electromagnetic imaging system based on multi-layer perception neural network and information metamaterials was constructed.

[0031] The specific steps of collecting the recognition imaging data of the observed target described in step 103 include:

[0032] Turn on the RF signal source, output a single-frequency RF signal, and irradiate it to the information metamaterial device after being radiated by the antenna. The information metamaterial device regulates the reflected electromagnetic wave. The reflected primary reflected electromagnetic wave irradiates the observation target and diffracts. The secondary reflected wave is collected by the receiving antenna in the far field area. The RF power meter records the signal power received by the receiving antenna. The coding combination of the information metasurface

[0033] The information metamaterial encoding FPGA device and the RF power meter are controlled by the host computer to work synchronously; the N groups of information codes of the information metasurface coding combination in step 102 are periodically read and cyclically loaded into the information metamaterial, so that in each clock cycle, the information metamaterial generates a set of different characteristic pattern radiation fields, and the RF power meter performs a data acquisition at the same time, and the acquired signal strength is recorded as y i After N sets of encoding cycles are completed, the intensity sequence obtained is

[0034] y=[y1 y2 … y N ] T (3).

[0035] The numerical model described in step 104, which is based on the electromagnetic field electromagnetic wave space propagation theory and describes the intrinsic relationship between the target electromagnetic information and the receiving end signal in the imaging system constructed in step 103, is specifically: under the i-th group of coding, the field intensity generated on the observed target by the primary reflected wave modulated by the information metamaterial reflection is w i , where w i is a K-dimensional column vector;

[0036] The electromagnetic information of the observed target object can be represented by a K-dimensional column vector x, where each element in x represents the quantized reflection information of the target object grid point to the electromagnetic wave;

[0037] According to the free propagation model of electromagnetic waves in space, the receiving end receives the signal under the i-th group of coding:

[0038]

[0039] where x j , w j are the jth elements of column vectors x and w respectively; U j is the propagation path coefficient from the jth pixel point to the observation antenna after the observation target position is quantized,

[0040] Ignoring the changes in the observed target under a set of periodic measurements, introducing space and link noise, and considering the direct path propagation between the excitation source antenna and the feedback antenna, the line-of-sight propagation signal needs to be added to the received signal. Therefore, the electromagnetic wave space propagation model of the imaging system can be expressed as

[0041] y=Wx+b+n (5).

[0042] The propagation path coefficient described above; since the receiving antenna is located in the far field area of ​​the observation surface, U j It has nothing to do with j, that is, the propagation path coefficient can be equivalent to 1.

[0043] The low-complexity linear algorithm for image reconstruction in step 105 is:

[0044] Based on the near-field quasi-linear relationship between the received signal strength sequence y and the observed target electromagnetic information x derived in step 104, a reconstruction algorithm is designed using a sample set driven method. The specific steps include:

[0045] Without considering the noise term, the imaging system electromagnetic wave space propagation model (5) is transformed into an equivalent form

[0046]

[0047] Select L groups of samples, collect data for each group of samples in step 103, and record the electromagnetic information of the target observed by the i-th group of samples as x i , the corresponding N-dimensional observation intensity sequence is y i

[0048] Combine the observed target electromagnetic information into matrix X

[0049]

[0050] Combine the observed intensities into a matrix Y

[0051] Y=[y1 y2 … y L ] (9)

[0052] Based on model formula (7), the relationship between the two is

[0053] Y=PX (10)

[0054] Where P is the actual transformation matrix. Considering the ideal characteristics such as the actual spatial propagation phase, it is somewhat different from the specific value in equation (7). The intensity b describing the line-of-sight propagation is unknown, so it is determined based on equation (10) using a sample-driven method. Since X and Y composed of L groups of sample data are known, the inverse transformation matrix Q used in the reconstruction algorithm is calculated as

[0055] Q=Xpinv(Y) (11)

[0056] Where pinv represents the Penrose inverse matrix, and Q is a (K+1)*N dimensional matrix. Take the first K rows and construct the inverse transformation matrix Based on a set of observation degree sequences y, the observed target information is inferred, denoted as The reconstruction algorithm is

[0057]

[0058] Beneficial effects: Compared with the prior art, the electromagnetic target recognition imaging method based on a multi-layer perception neural network and metamaterials provided by the present invention has the following beneficial effects:

[0059] 1. With the help of MLP network, the feature space of the observed target sample set is extracted, and with the help of information metamaterials, real-time projection is performed in the electromagnetic space domain. The system has significant target recognition accuracy, image reconstruction accuracy, and improved noise interference resistance.

[0060] 2. With the help of the MLP network feature extraction results, the number of measurements in the imaging task is far less than the number of pixels of the reconstructed image, with a high compression rate, the imaging rate is greatly improved, and it has real-time imaging capabilities.

[0061] 4. The system autonomously optimizes the sample set and designs the coding combination in a targeted manner. The system is adaptive and suitable for imaging recognition tasks under various environmental targets.

[0062] 5. The system is based on single frequency point measurement. The antenna transmitting port does not need additional signal modulation, and the receiving port only needs power detection. The system receiving and transmitting structure is greatly simplified. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments.

[0064] Figure 1 It is a flow chart of an electromagnetic target recognition and imaging method based on a multi-layer perception neural network and information metamaterials provided by an embodiment of the present invention;

[0065] Figure 2 is a schematic diagram of the MLP network structure constructed in step 101 and a schematic diagram of feature extraction and encoding generation;

[0066] Figure 3 It is a schematic diagram of the physical system structure of an electromagnetic target recognition and imaging method based on a multi-layer perception neural network and information metamaterials provided by an embodiment of the present invention;

[0067] In the figure, there are: 301-transmitting antenna, 302-information metamaterial device, 303-observation target, 304-information metamaterial coding control FPGA device, 305-receiving antenna, 306-RF power meter, 307-computer;

[0068] Figure 4 It is a schematic diagram of image recognition results under Case A of the electromagnetic target recognition imaging method based on a multi-layer perception neural network and information metamaterials provided by the present invention. Figure 4 (a) is the trend curve of recognition accuracy under different noise powers. Figure 4 (b) is the confusion matrix of the classification prediction results for 10 types of targets in the absence of noise.

[0069] Figure 5It is a schematic diagram of image reconstruction results under Case A of the electromagnetic target recognition imaging method based on multi-layer perception neural network and information metamaterial provided by the present invention.

[0070] Figure 6 This is a schematic diagram of the image reconstruction resolution results of the electromagnetic target recognition imaging method based on the multi-layer perception neural network and information metamaterial provided by the present invention under Case B. DETAILED DESCRIPTION

[0071] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application. The present invention is further explained below in conjunction with the accompanying drawings.

[0072] The present invention discloses an electromagnetic target recognition imaging method based on a multi-layer perceptual neural network and information metamaterials. Figure 1 As shown, the following steps are included:

[0073] Step 101: Based on the imaging target potential sample set data, a multi-layer perceptron neural network (MLP) is constructed and trained on the selected sample set. As an optional case A, the handwritten character patterns in the MNIST data set are used as the imaging target sample set. The constructed MLP network has 4 layers, including an input layer and an output layer. The number of neurons in the 4 layers is 400, 64, 64, and 10 respectively. The network is trained using the back propagation algorithm. The pattern feature space is then extracted from the trained multi-layer perceptron neural network. In the MLP network model, the relationship between the neurons in the i-th layer and the neurons in the i+1-th layer is

[0074] a i+1 =F(W i ·a i ) i=1,2,…,M-1 (1)

[0075] Where M = 4 is the number of MLP network layers, a i , a i+1 are the neuron outputs of the i-th layer and the i+1-th layer, respectively, W i is the weight matrix between the i-th layer and the i+1-th layer, and F() is the nonlinear activation function.

[0076] The MLP network established for feature extraction of latent sample set data of imaging targets has the same number of input neurons as the number of pixels of the imaging target, and the number of output neurons is determined by the specific type of the network. If the ultimate goal of the imaging system is to identify the target, the number of output neurons is the number of corresponding label categories. If the ultimate goal of the imaging system is to image the target, the number of output neurons should be the same as the number of pixels of the imaging target. In any case, the number of neurons in each initial layer of the network should decrease layer by layer to achieve feature compression. In this case, the MLP network is selected as the classification form, so the number of output neurons is 10, which is consistent with the total number of categories in the sample set.

[0077] In the network trained by the optimization algorithm, the adjustable parameter W in the network i All have been optimized to reflect the feature space of the sample set. In particular, the weight matrix W1 between the first and second layers is the selected main feature space matrix.

[0078] Step 102: Based on the feature space, perform holographic projection inverse reconstruction on the feature vectors or patterns one by one, so that the primary reflected wavefront reflected by the feed source after passing through the information metasurface has a characteristic pattern field intensity distribution, thereby calculating and generating a set of coding combinations of the information metasurface. Figure 3 As shown, the feature space matrix W1 extracted in step 101 can be expressed as

[0079] W1=[w1 w2 … w N ] T (2)

[0080] Where N = 64 positive integers, representing the number of rows of matrix W1, that is, the dimension of the feature space; T represents the transposition symbol; w1 to w 64 There are 64 sets of feature column vectors. The length of the column vectors is K=400, where K is a positive integer representing the number of pixels of the imaging target.

[0081] Each set of feature vectors w i The image is converted into a 20×20 image matrix with the same size as the observed target. Based on its numerical intensity data, the information metasurface encoding is calculated using the holographic projection inverse reconstruction modified Gerchberg-Saxton (GS) algorithm, so that the primary reflected wavefront reflected by the feed source after passing through the information metasurface has a characteristic pattern field w i The N=64 sets of codes obtained by calculation are stored in FPGA for subsequent calls.

[0082] Figure 2 (a) shows a schematic diagram of an MLP network with 4 layers of neurons. Figure 2 (b)(d) show the results of converting the first and second set of eigenvectors w_1 in Case A to image matrices respectively. Figure 2 (c) and (e) show the generation of Figure 2One-bit metasurface coding designed with the near-field characteristic patterns shown in (b) and (d).

[0083] Step 103: Establish an imaging system consisting of a feed antenna, an information metamaterial, an observation object platform, and a receiving antenna. Figure 2 As shown, the specific steps include:

[0084] Based on the position of the observation target 303, the metamaterial device 302 is placed in front of the observation target. The sample target of the MNIST data set in the case is made of conductive ink, and the effective range of the observation surface is 280mm×280mm. The metamaterial device selects a reflective single-bit phase-controlled metasurface, the array unit size is 30×30, and the system center frequency is 9.0GHz.

[0085] Based on the relative position of the observed target and the metamaterial device, the excitation source antenna 301 is installed to ensure that the excitation source center and the observed target are mirror-symmetric about the metamaterial device, and the main radiation direction of the excitation source antenna is aligned with the center of the metamaterial device, and the beam can cover the metamaterial device.

[0086] Based on the size and position of the observation target, the receiving antenna 305 is installed to ensure that the main radiation direction of the receiving antenna is aimed at the observation target, and the receiving antenna is located in a relatively far field area of ​​the observation target.

[0087] Connect the excitation source antenna to the RF signal source generator, and connect the receiving antenna to the RF power meter 306. The metamaterial device is connected to the FPGA coding control device 304. Connect the host computer 307 to the RF power meter and the coding control FPGA device respectively. In this case, the RF signal source generator and the RF power meter are jointly implemented by an Agilent Formula N5230C vector network analyzer.

[0088] In this way, an intelligent electromagnetic imaging system based on multi-layer perception neural network and information metamaterials was constructed.

[0089] The imaging recognition data collection process described in step 103 specifically includes the following steps:

[0090] Turn on the RF signal source and output a single-frequency RF signal. After being radiated by the antenna, it is irradiated to the information metamaterial device. The information metamaterial device regulates the reflected electromagnetic wave. The reflected primary reflected electromagnetic wave irradiates the observation target and diffracts. The secondary reflected wave is collected by the receiving antenna in the far field area. The RF power meter records the signal power received by the receiving antenna.

[0091] The host computer controls the information metamaterial encoding FPGA device and the RF power meter to work synchronously. From the N=64 sets of information encoding generated in step 102, the periodic reading cycle is loaded into the information metamaterial, so that in each clock cycle, the information metamaterial generates a set of different characteristic pattern radiation w i At the same time, the RF power meter collects data once, and the collected signal strength is recorded as y i After N sets of encoding cycles are completed, the intensity sequence obtained is

[0092] y=[y1 y2 … y N ] T (3)

[0093] Step 104: Establish an electromagnetic wave space propagation model for the imaging system based on electromagnetic field theory. Under the i-th group of coding, the field intensity generated by the primary reflected wave modulated by the information metamaterial reflection on the observation target is w i , where w i is a K=400-dimensional column vector.

[0094] The electromagnetic information of the observed target object can be represented by a K-dimensional column vector x. Each element in x represents the quantized reflection information of the target object grid point to the electromagnetic wave.

[0095] According to the free propagation model of electromagnetic waves in space, the receiving end receives the signal under the i-th group of coding:

[0096]

[0097] where x j , w j are the jth elements of column vectors x and w respectively. j is the propagation path coefficient from the jth pixel point to the observation antenna after the observation target position is quantized. In particular, since the receiving antenna is located in the far field area of ​​the observation surface, U j It has nothing to do with j, that is, the propagation path coefficient can be equivalent to 1.

[0098] Ignore the changes of the observed target under a set of periodic measurements, introduce space and link noise. Considering the direct path propagation between the excitation source antenna and the feedback antenna, the line-of-sight propagation signal needs to be added to the received signal, so the electromagnetic wave space propagation model of the imaging system can be expressed as

[0099] y=Wx+b+n (5)

[0100] Step 105: Based on the intrinsic relationship between the received electromagnetic wave signal and the electromagnetic information of the observed target in the model derived in step 104, a low-complexity linear algorithm for image reconstruction is designed. Based on the near-field quasi-linear relationship between the received signal strength sequence y and the electromagnetic information x of the observed target derived in step 104, a reconstruction algorithm is designed using a sample set driven method. The specific steps include:

[0101] Without considering the noise term, the imaging system electromagnetic wave space propagation model (5) is transformed into an equivalent form

[0102]

[0103] Select L=1000 groups of samples, and collect data for each group of samples in step 103. Let the electromagnetic information of the target observed by the i-th group of samples be x i , the corresponding N-dimensional observation intensity sequence is y i .

[0104] Combine the observed target electromagnetic information into matrix X

[0105]

[0106] Combine the observed intensities into a matrix Y

[0107] Y=[y1 y2 … y L ] (9)

[0108] Based on model formula (7), the relationship between the two is

[0109] Y=PX (10)

[0110] Where P is the actual transformation matrix. Considering the ideal characteristics such as the actual spatial propagation phase, it is somewhat different from the specific value in formula (7). The intensity b describing the line-of-sight propagation is unknown, so it is determined by sample driving based on formula (10). Since X and Y composed of L groups of sample data are known, the inverse transformation matrix Q used in the reconstruction algorithm is calculated as

[0111] Q=Xpinv(Y) (11)

[0112] Where pinv represents the Penrose inverse matrix, and Q is a (K+1)*N dimensional matrix. Take the first K rows and construct the inverse transformation matrix Based on a set of observation degree sequences y, the observed target information is inferred, denoted as The reconstruction algorithm is

[0113]

[0114] Figure 4 This is the imaging target recognition result of case A using the MNIST sample set as the target. Figure 4As shown in (a), the confusion matrix shows the recognition accuracy of each category under 1000 groups of samples. The accuracy of the whole sample set reaches 77.7%, which has a high recognition accuracy. In order to compare the advantages of spatial domain feature extraction based on MLP network, we chose to replace the MLP neural network guided information metamaterial encoding method with random encoding. Similarly, 64 groups of random encoding were selected to measure the single frequency point of the whole sample set. The accuracy of the two methods changes with the injected noise signal power as shown in Figure 4 As shown in (b), the MLP neural network-based imaging recognition method has a maximum accuracy improvement of 18% compared to the traditional method, proving that the system has a certain ability to resist noise interference.

[0115] Figure 5 This is the result of imaging target recognition in case A with the MNIST sample set as the target. The 10 images are reconstructed images of a group of samples randomly selected from 10 types of samples, and the image size is 20*20 pixels. In order to quantify the reconstruction effect, the mean square error (MSE) is introduced to measure the reconstruction error, which is defined as

[0116]

[0117] Where x is the column vector of the original image after stretching, is the column vector of the reconstructed image after stretching, and L = 400 is the total number of pixels. Table 1 shows the average MSE of the reconstructed image of the whole sample set under the MLP network assisted coding design, where four noise powers are selected to simulate the noise environment. According to the statistical data in Table 1, the MSE of the reconstructed image loss decreased by 10% on average, proving the effectiveness of the imaging system based on the MLP network and information metamaterials.

[0118] Table 1

[0119] Normalized noise power 0 0.01 0.05 0.1 MLP method average MSE 0.1632 0.1696 0.211 0.1991 Average MSE of random methods 0.1565 0.1656 0.1799 0.1858

[0120] Figure 6 The image target recognition result of case B with metal stripe sample set as the target. In case B, the selected imaging target sample set is a 27mm wide metal strip structure, where the sample set consists of one or two strips at different horizontal positions on the 270mm×270mm observation surface. Figure 6 (a)(b)(c) show the imaging reconstruction results of a metal strip target. The shaded part is the theoretical result of the metal strip target, and the actual part is the reconstruction result. It can be observed that the reconstruction result can correctly correspond to the target position and the theoretical width of the metal strip. Figure 6 (d)-(i) show the reconstruction results of the dual metal stripe target. Figure 6In the results (g)(h)(i), the distance between the two metals is only 13 mm, which is one-fourth of the diffraction wavelength of the observation surface. At this time, the imaging system can still correctly distinguish the relative position and gap between the two, verifying the high-resolution characteristics of the imaging system.

Claims

1. An electromagnetic target recognition imaging method based on a multi-layer perceptual neural network and metamaterials, characterized in that: The steps include: Step 101: Based on the imaging target potential sample set data, a multi-layer perceptron neural network MLP is constructed and trained on the selected sample set, and then a pattern feature space is extracted from the trained multi-layer perceptron neural network; Step 102: Based on the feature space, holographic projection inverse reconstruction is performed on the feature vectors or patterns one by one, so that the primary reflected wavefront reflected by the feed source after passing through the information metasurface has a characteristic pattern field intensity distribution, thereby calculating and generating a set of coding combinations of the information metasurface; Step 103: establishing an imaging system consisting of a feed antenna, an information metamaterial, an observation object platform, and a receiving antenna device, and collecting identification imaging data of the observation target; Step 104: establishing a numerical model describing the intrinsic relationship between the target electromagnetic information and the receiving end signal in the imaging system constructed in step 103 based on the electromagnetic field electromagnetic wave space propagation theory; Step 105: Based on the intrinsic relationship between the received electromagnetic wave signal and the electromagnetic information of the observed target in the numerical model derived in step 104, a low-complexity linear algorithm for image reconstruction is designed; in, The numerical model described in step 104, which is based on the electromagnetic field electromagnetic wave space propagation theory and describes the intrinsic relationship between the target electromagnetic information and the receiving end signal in the imaging system constructed in step 103, is specifically: under the i-th group of coding, the field intensity generated on the observed target by the primary reflected wave modulated by the information metamaterial reflection is w i , where w i is a K-dimensional column vector; The electromagnetic information of the observed target object can be represented by a K-dimensional column vector x, where each element in x represents the quantized reflection information of the target object grid point to the electromagnetic wave; According to the free propagation model of electromagnetic waves in space, the receiving end receives the signal under the i-th group of coding: where x j , w j are the jth elements of column vectors x and w respectively; U j is the propagation path coefficient from the jth pixel point to the observation antenna after the observation target position is quantized, Ignoring the changes of the observed target under a set of periodic measurements, introducing space and link noise, and considering the direct path propagation between the excitation source antenna and the feedback antenna, the line-of-sight propagation signal needs to be added to the received signal. Therefore, the electromagnetic wave space propagation model of the imaging system can be expressed as y=Wx+b+n(5); The low-complexity linear algorithm for image reconstruction in step 105 is: Based on the near-field quasi-linear relationship between the received signal strength sequence y and the observed target electromagnetic information x derived in step 104, a reconstruction algorithm is designed using a sample set driven method. The specific steps include: Without considering the noise term, the imaging system electromagnetic wave space propagation model (5) is transformed into an equivalent form Select L groups of samples, collect data for each group of samples in step 103, and record the electromagnetic information of the target observed by the i-th group of samples as x i , the corresponding N-dimensional observation intensity sequence is y i Combine the observed target electromagnetic information into matrix X Combine the observed intensities into a matrix Y Y=[y1 y2 … y L ] (9) Based on the model formula (7), the relationship between the two is Y = PX (10) Where P is the actual transformation matrix. Considering the ideal characteristics of the actual spatial propagation phase, it is somewhat different from the specific value in equation (7). The intensity b describing the line-of-sight propagation is unknown, so it is determined based on equation (10) using a sample-driven method. Since X and Y composed of L groups of sample data are known, the inverse transformation matrix Q used in the reconstruction algorithm is calculated as Q = Xpinv(Y) (11) Where pinv represents the Penrose inverse matrix, Q is a (K+1)*N dimensional matrix; take the first K rows and construct the inverse transformation matrix Based on a set of observation degree sequences y, the observed target information is inferred, denoted as The reconstruction algorithm is 2. The electromagnetic target recognition imaging method based on multi-layer perceptual neural network and metamaterial according to claim 1 is characterized in that: The imaging target potential sample set described in step 101 is extracted based on the feature space of the multi-layer perceptron neural network MLP; in the MLP network model, the relationship between the i-th layer neuron and the i+1-th layer neuron is a i+1 =F(W i ·a i ) i=1,2,…,M-1(1) Where M is the number of MLP network layers, a i , a i+1 are the neuron outputs of the i-th layer and the i+1-th layer, respectively, W i is the weight matrix between the i-th layer and the i+1-th layer, and F() is the nonlinear activation function; The MLP network established for feature extraction of latent sample set data of imaging targets has the same number of input neurons as the number of pixels of the imaging target, and the number of output neurons is determined by the specific type of network. In any case, the number of neurons in each initial layer of the network should decrease layer by layer to achieve feature compression. In the network trained by the optimization algorithm, the weight matrix W i All have been optimized to reflect the feature space of the sample set.

3. The electromagnetic target recognition imaging method based on multi-layer perceptual neural network and metamaterial according to claim 2 is characterized in that The number of output neurons is determined by the specific type of network. When the ultimate goal of the imaging system is to identify the target, the number of output neurons is the corresponding number of label categories; when the ultimate goal of the imaging system is to image the target, the number of output neurons should be the same as the number of pixels of the imaging target.

4. The electromagnetic target recognition imaging method based on multi-layer perceptual neural network and metamaterial according to claim 2 is characterized in that The weight matrix W i , the weight matrix W1 between the first layer and the second layer is the selected main feature space matrix; the main feature space matrix W1 can be expressed as W1 = [w1 w2 … w N ] T (2) Where N is a positive integer, representing the number of rows of matrix W1, that is, the dimension of the feature space; T represents the transposition symbol; w1 to w N are N groups of feature column vectors; the length of the column vectors is K, where K is a positive integer representing the number of pixels of the imaging target.

5. The electromagnetic target recognition imaging method based on multi-layer perceptual neural network and metamaterial according to claim 1 is characterized in that: In step 102, the holographic projection inverse reconstruction is performed on the feature vectors or patterns one by one; Each set of feature vectors w constructed in step 101 i The image matrix is ​​converted into the same size as the observed target. Based on its numerical intensity data, the information metasurface encoding is calculated using the holographic projection inverse reconstruction modified Gerchberg-Saxton algorithm, so that the primary reflected wavefront reflected by the feed source after passing through the information metasurface has a characteristic vector w i The distribution of the converted feature pattern field and the N groups of codes obtained by calculation are stored in the FPGA for subsequent calls.

6. The electromagnetic target recognition imaging method based on multi-layer perceptual neural network and metamaterial according to claim 1 is characterized in that: The step 103 of establishing an imaging system consisting of a feed antenna, an information metamaterial, an observation object platform, and a receiving antenna device comprises the following specific steps: Based on the position of the observation target (303), the metamaterial device (302) is placed in front of the observation target, Based on the relative position between the observation target position and the metamaterial device, the excitation source antenna (301) is installed to ensure that the excitation source center and the observation target are mirror-symmetrical about the metamaterial device, and the main radiation direction of the excitation source antenna is aligned with the center of the metamaterial device, and the beam can cover the metamaterial device; Based on the size and position of the observation target, a receiving antenna (305) is installed to ensure that the main radiation direction of the receiving antenna is aligned with the observation target and the receiving antenna is located in a relatively far field area of ​​the observation target; Connecting the excitation source antenna to the radio frequency signal source generator, and connecting the receiving antenna to the radio frequency power meter (306); connecting the metamaterial device to the FPGA coding control device (304); connecting the host computer (307) to the radio frequency power meter and the coding control FPGA device respectively; In this way, an intelligent electromagnetic imaging system based on multi-layer perception neural network and information metamaterials was constructed.

7. The electromagnetic target recognition imaging method based on multi-layer perceptual neural network and metamaterial according to claim 1 is characterized in that: The specific steps of collecting the recognition imaging data of the observed target described in step 103 include: Turn on the RF signal source, output a single-frequency RF signal, and irradiate it to the information metamaterial device after being radiated by the antenna. The information metamaterial device regulates the reflected electromagnetic wave. The reflected primary reflected electromagnetic wave irradiates the observation target and diffracts. The secondary reflected wave is collected by the receiving antenna in the far field area. The RF power meter records the signal power received by the receiving antenna. The coding combination of the information metasurface; The information metamaterial encoding FPGA device and the RF power meter are controlled by the host computer to work synchronously; the N groups of information codes of the information metasurface coding combination in step 102 are periodically read and cyclically loaded into the information metamaterial, so that in each clock cycle, the information metamaterial generates a set of different characteristic pattern radiation fields, and the RF power meter performs a data acquisition at the same time, and the acquired signal strength is recorded as y i After N sets of encoding cycles are completed, the intensity sequence obtained is y=[y1 y2 … y N ] T (3)。 8. The electromagnetic target recognition imaging method based on multi-layer perceptual neural network and metamaterial according to claim 1 is characterized in that: The propagation path coefficient described above; since the receiving antenna is located in the far field area of ​​the observation surface, U j It is independent of j, that is, the propagation path coefficient is equivalent to 1.

Citation Information

Patent Citations

  • Multispectral single-pixel imaging deep learning image reconstruction method

    CN110175971A

  • Target detection method based on remote sensing electromagnetic radiation and infrared image

    CN113723511A

  • Electromagnetic wave space calculation apparatus and operation method therefor

    WO2021223512A1