Back-wall target two-dimensional fast imaging method based on two-way complex convolutional neural network

Abstract

The invention discloses a two-dimensional fast imaging method for a target behind a wall based on a double-path complex convolutional neural network, which belongs to the technical field of fast imaging and comprises the following steps of: transmitting and receiving electromagnetic waves by using a multi-antenna array, measuring S parameters among multiple antennas by using a vector network analyzer, processing the S parameters by using the double-path complex convolutional neural network, and obtaining the target after the wall based on the double-path complex convolutional neural network. And obtaining an after-wall target two-dimensional imaging result. According to the complex convolutional neural network through-wall imaging method provided by the invention, training is carried out through a large amount of actually measured data, the model robustness is high, a complex scene can be dealt with, a graphics processor acceleration technology of a computer is fully utilized, the imaging speed is high, and real-time imaging can be realized; according to the complex convolutional neural network through-wall imaging method, a double-path design is adopted, an imaging optimization process is divided into one-dimensional range profile optimization and two-dimensional horizontal section imaging optimization, an after-wall target imaging result can be effectively optimized, and a user can conveniently observe needed information from the after-wall target imaging result.

Description

Technical Field

[0001] The invention discloses a two-dimensional rapid imaging method for a target behind a wall based on a dual-path complex convolutional neural network, belonging to the technical field of rapid imaging. Background Art

[0002] Microwave through-wall imaging technology uses low-frequency electromagnetic waves emitted through a wall to sense, locate, and image targets behind it based on their echo signals. Microwave through-wall imaging technology boasts excellent penetration and is not easily affected by weather, light, or temperature. The signals emitted can be classified into three types: pulsed, step-frequency continuous wave, and frequency-modulated continuous wave. Step-frequency continuous wave microwave signals offer relatively flexible operating frequencies, strong interference resistance, a wide dynamic range, and improved detection performance. Traditional stepped-frequency microwave imaging methods typically use a signal source connected to an external antenna to transmit microwave signals of varying frequencies through a wall. The signals penetrate the wall, reflect off the target, and then pass through the wall a second time before being received by a receiving antenna. The amplitude ratio and phase difference between the transmitted and received signals are then measured, resulting in a high signal-to-noise ratio signal using moving target indication technology. This allows the calculation of a one-dimensional range profile of the target. Multiple antennas are arranged in a linear or planar array, and multiple one-dimensional range profiles are superimposed through backprojection to create a two-dimensional or three-dimensional image. The use of deep learning methods in microwave fluoroscopy can adapt to varying wall materials and environmental variations, improving imaging accuracy and applicability. In practical applications, directly performing 3D imaging of the entire detection area is computationally extremely complex. 2D imaging can quickly determine the position of people behind walls and also partition local spaces for 3D imaging.

[0003] Microwave signals experience significant refraction, multipath reflection, and attenuation when penetrating walls. Traditional through-wall radar imaging algorithms use complex mathematical models and extensive computing resources to compensate for microwave signals that penetrate walls. This results in slow signal processing, making real-time imaging difficult. Furthermore, in complex scenarios, the signal-to-noise ratio of the echo signal from targets behind the wall is low, making traditional through-wall radar imaging algorithms prone to failure. Existing methods using deep learning models for microwave through-wall imaging generally optimize traditional imaging results through deep learning. Specifically, imaging is first performed using a traditional algorithm, and then the resulting image is fed into a deep learning model for optimization. However, traditional algorithms require compensation for signals that penetrate walls, making it difficult to accurately estimate the physical parameters of walls. Furthermore, the scenes behind walls are complex and diverse, and the resulting images generated by traditional algorithms are prone to losing information about the targets behind the wall. This results in poor data integrity and cannot guarantee the authenticity of the data input to the deep learning model. Furthermore, this computationally intensive stage prevents end-to-end deep learning from being a viable solution, hindering real-time imaging. Summary of the Invention

[0004] The purpose of the present invention is to provide a two-dimensional rapid imaging method for targets behind a wall based on a dual-path complex convolutional neural network, so as to solve the problems in the prior art that the traditional microwave through-wall imaging method has low imaging resolution, is relatively difficult for users to observe effective information, and is prone to failure in complex scenarios.

[0005] A two-dimensional fast imaging method for targets behind a wall based on a dual-path complex convolutional neural network includes using a multi-antenna array to transmit and receive electromagnetic waves, using a vector network analyzer to measure the S parameters between the multiple antennas, and using a dual-path complex convolutional neural network to process the S parameters to obtain a two-dimensional imaging result of the target behind the wall.

[0006] The two-way complex convolutional neural network consists of two branches.

[0007] The first branch inputs the S parameter of size T×C×P, performs moving target display, obtains the first data of size T×C×N×P, and then performs one-dimensional complex convolution to obtain the first data of size The second data is then convolved to obtain a one-dimensional complex convolution of size The third data.

[0008] In the second branch, the first data is spatially projected to obtain fourth data, the fourth data is subjected to two-dimensional complex convolution to obtain fifth data, the second data is spatially projected and combined with the fifth data to perform two-dimensional complex convolution to obtain sixth data, and the third data is spatially projected and combined with the sixth data to obtain a position prediction result.

[0009] Complex convolution involves treating complex numbers as two-dimensional real number pairs, denoted by z = A + iB, where z is a complex number, i is a complex unit, A and b are two-dimensional real number pairs of complex numbers. The essence of the convolution operation is the product and accumulation of the feature matrix and the weight matrix. The complex weight is denoted by W and the complex eigenvalue is h:

[0010] W=A+iB;

[0011] h=x+iy;

[0012] Where x and y are two-dimensional real number pairs corresponding to complex eigenvalues, and the convolution operation on h using W is:

[0013] W*h=(A*xB*y)+i(B*x+A*y).

[0014] Compared with the existing technology, the present invention has the following beneficial effects: the complex convolutional neural network through-wall imaging method provided by the present invention is trained through a large amount of measured data, the model has strong robustness, can cope with complex scenes, and makes full use of the computer's graphics processor acceleration technology, with fast imaging speed and real-time imaging; the complex convolutional neural network through-wall imaging method adopts a dual-path design, splitting the imaging optimization process into one-dimensional range image optimization and two-dimensional horizontal section imaging optimization, which can effectively optimize the imaging results of the target behind the wall, making it easier for users to observe the required information from it; the complex convolutional neural network through-wall imaging method is an end-to-end neural network model, which is not limited by traditional imaging methods and can directly extract features and image from the collected stepped frequency original signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of area array stepped frequency imaging;

[0016] Figure 2 This is a two-way complex convolutional neural network model diagram;

[0017] Figure 3 is the target's real location map;

[0018] Figure 4 It is a traditional BP imaging image;

[0019] Figure 5 It is a one-dimensional complex feature space projection plus visualization graph;

[0020] Figure 6 This is the final prediction result diagram of the two-dimensional complex feature. DETAILED DESCRIPTION

[0021] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0022] A two-dimensional fast imaging method for targets behind a wall based on a dual-path complex convolutional neural network includes using a multi-antenna array to transmit and receive electromagnetic waves, using a vector network analyzer to measure the S parameters between the multiple antennas, and using a dual-path complex convolutional neural network to process the S parameters to obtain a two-dimensional imaging result of the target behind the wall.

[0023] The two-way complex convolutional neural network consists of two branches.

[0024] The first branch inputs the S parameter of size T×C×P, performs moving target display, obtains the first data of size T×C×N×P, and then performs one-dimensional complex convolution to obtain the first data of size The second data is then convolved to obtain a one-dimensional complex convolution of size The third data.

[0025] In the second branch, the first data is spatially projected to obtain fourth data, the fourth data is subjected to two-dimensional complex convolution to obtain fifth data, the second data is spatially projected and combined with the fifth data to perform two-dimensional complex convolution to obtain sixth data, and the third data is spatially projected and combined with the sixth data to obtain a position prediction result.

[0026] Complex convolution involves treating complex numbers as two-dimensional real number pairs, denoted by z = A + iB, where z is a complex number, i is a complex unit, A and b are two-dimensional real number pairs of complex numbers. The essence of the convolution operation is the product and accumulation of the feature matrix and the weight matrix. The complex weight is denoted by W and the complex eigenvalue is h:

[0027] W=A+iB;

[0028] h=x+iy;

[0029] Where x and y are two-dimensional real number pairs corresponding to complex eigenvalues, and the convolution operation on h using W is:

[0030] W*h=(A*xB*y)+i(B*x+A*y).

[0031] The present invention addresses the problems of low imaging resolution of traditional microwave through-wall imaging methods, the relative difficulty for users to observe effective information from them, and the tendency to fail in complex scenarios. It adopts stepped-frequency continuous waves for microwave perspective imaging and adopts deep learning methods for imaging to adapt to different wall materials and environmental changes, thereby improving the accuracy and applicability of imaging. In response to the problem that existing deep learning methods cannot be directly applied to the raw data of stepped-frequency signals, a method for imaging targets behind walls based on complex convolutional neural networks is proposed, and a complex neural network model that can directly process the raw data is designed to achieve end-to-end model training and imaging prediction. In response to the problems of high computational complexity and high video memory usage of complex convolutional neural networks, a dual-path complex convolutional neural network is designed. One path efficiently extracts one-dimensional features, and the other path performs complex convolution and position prediction on the two-dimensional features after spatial projection. This method has fast positioning speed and low video memory usage.

[0032] In traditional imaging principles, the formula for step frequency transmission signal is:

[0033] S k =A i exp[-j2π(f0+kΔf)];

[0034] Schematic diagram of area array stepped frequency imaging Figure 1 As shown in the figure, the stepped frequency signal is transmitted by M antennas and received by N antennas. The distance between antenna m, n and the target is R m 、R n , then the target's echo signal is:

[0035]

[0036] The phase of the echo signal is related to the target distance. The target can be imaged by calculating the distance from different positions in space to the antenna and performing distance compensation and superposition on the signals. The calculation formula of the direct frequency domain projection algorithm is as follows:

[0037]

[0038] In the direct frequency domain projection algorithm, the echo signal can be viewed as a complex feature, and the phase compensation can be viewed as a complex weight. Therefore, complex convolution can completely replace the direct frequency domain projection algorithm, providing a theoretical basis for implementing end-to-end deep learning models. Therefore, to ensure the authenticity and validity of the input data and realize end-to-end microwave radioscopy imaging based on deep learning, we take advantage of the characteristics of frequency domain data collected by the vector network analyzer and design a complex neural network to directly process the S parameters collected by the vector network analyzer. This effectively preserves the amplitude and phase information of the frequency domain signal and maximizes the use of the reflection information of the target behind the wall. This allows the complex neural network to better extract the reflection characteristics of the target behind the wall from the S parameters, improving the imaging effect.

[0039] The computational complexity of complex convolution is four times that of traditional convolution, and when the product of the number of transmitting and receiving antennas and the number of scanning points of the stepped frequency signal is high, the computational complexity is too large. Therefore, a dual-path complex convolutional neural network model is proposed, such as Figure 2 As shown, each signal is processed separately through one-dimensional complex convolution to extract one-dimensional complex features. These features are then projected onto a two-dimensional space and subjected to two-dimensional complex convolution, thereby optimizing computational complexity and improving imaging speed. Visual analysis of the one-dimensional information extracted by the proposed method reveals higher coherence across the optimized channels, and a clear focus after projection and superposition. This method offers greater interpretability than methods that first image and then detect using a deep learning model.

[0040] The actual position of the target in the present invention is as follows Figure 3 As shown, BP imaging is Figure 4If BP imaging is used directly, the one-dimensional distance signals of different channels need to be projected into the same two-dimensional space for superposition. When the noise is large or the target movement is complex, the image superimposed by multiple signals is scattered, and the target is hidden in the interference signals such as coupling waves and noise, making it difficult to distinguish the target. Using one-dimensional complex convolution, as shown Figure 5 As shown in the figure, the human target features are directly extracted from the measured S parameters, and the channels of the hidden layer are directly superimposed to visualize the hidden channels. The one-dimensional information extracted by the proposed method is visualized and analyzed to obtain the final prediction results of the two-dimensional complex features as shown in the figure. Figure 6 As shown in the figure, the coherence of each channel after optimization is higher, and the focus position is obvious after projection and superposition. It is more interpretable than the method of first imaging and then detecting with a deep learning model.

[0041] The spatially projected signals are superimposed and further extracted by the complex neural network to obtain the final output result. The label position, the imaging results of the traditional algorithm, and the hidden channel feature map extracted by the convolutional neural network are visualized and compared with the final output below. It can be seen that the proposed algorithm can accurately predict the target position.

[0042] The present invention is a method for imaging targets behind walls by performing end-to-end feature extraction on stepped frequency signals based on a complex convolutional neural network, making full use of the amplitude and phase information of the echo signal. A dual-path complex convolutional neural network is designed, where one path efficiently extracts one-dimensional features, and the other path performs complex convolution and position prediction on the two-dimensional features after spatial projection, thereby reducing the memory usage and computational complexity of the complex neural network and speeding up the imaging speed. An end-to-end complex neural network model is designed for vector network acquisition data, effectively retaining the amplitude and phase information of the stepped frequency echo signal, with high credibility and strong interpretability. The complex convolutional neural network adopts a dual-path model design, where one path efficiently extracts one-dimensional features, and the other path performs complex convolution and position prediction on the two-dimensional features after spatial projection, resulting in fast positioning speed and low memory usage.

[0043] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A two-dimensional rapid imaging method for targets behind a wall based on a dual-path complex convolutional neural network, characterized in that: It includes using a multi-antenna array to transmit and receive electromagnetic waves, using a vector network analyzer to measure the S parameters between multiple antennas, and using a dual-path complex convolutional neural network to process the S parameters to obtain two-dimensional imaging results of the target behind the wall.

2. The method for rapid two-dimensional imaging of a target behind a wall based on a dual-path complex convolutional neural network according to claim 1, characterized in that: The two-way complex convolutional neural network consists of two branches.

3. The method for rapid two-dimensional imaging of a target behind a wall based on a dual-path complex convolutional neural network according to claim 2, characterized in that: The first branch inputs the S parameter of size T×C×P, performs moving target display, obtains the first data of size T×C×N×P, and then performs one-dimensional complex convolution to obtain the first data of size The second data is then convolved to obtain a one-dimensional complex convolution of size The third data.

4. The method for rapid two-dimensional imaging of a target behind a wall based on a dual-path complex convolutional neural network according to claim 3, characterized in that: In the second branch, the first data is spatially projected to obtain fourth data, the fourth data is subjected to two-dimensional complex convolution to obtain fifth data, the second data is spatially projected and combined with the fifth data to perform two-dimensional complex convolution to obtain sixth data, and the third data is spatially projected and combined with the sixth data to obtain a position prediction result.

5. The method for rapid two-dimensional imaging of a target behind a wall based on a dual-path complex convolutional neural network according to claim 4, characterized in that: Complex convolution involves treating complex numbers as two-dimensional real number pairs, denoted by z = A + iB, where z is a complex number, i is a complex unit, A and b are two-dimensional real number pairs of complex numbers. The essence of the convolution operation is the product and accumulation of the feature matrix and the weight matrix. The complex weight is denoted by W and the complex eigenvalue is h: W=A+iB; h=x+iy; Where x and y are two-dimensional real number pairs corresponding to complex eigenvalues, and the convolution operation on h using W is: W*h=(A*xB*y)+i(B*x+A*y).

Images