Terahertz laminated imaging system based on quantum cascade laser and physical driving neural network
By using quantum cascade lasers and physically driven neural networks in terahertz stacked imaging technology, the problems of poor portability and insufficient reconstruction effect in the prior art are solved, and efficient terahertz wave utilization and high-quality imaging effects are achieved.
Patent Information
- Application Number
- CN202411645777.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-18
- Publication Date
- 2025-03-04
AI Technical Summary
The existing terahertz stacked imaging technology has problems such as poor portability, high light source dependence and low signal-to-noise ratio in experimental design and reconstruction algorithms, which limits its practical application.
The terahertz stacked imaging system based on quantum cascade laser and physically driven neural network is adopted. The quantum cascade laser is used as a compact light source and is reconstructed in combination with the physically driven neural network to improve the portability of the system and the robustness of the reconstruction effect.
It realizes efficient terahertz wave utilization, improves the robustness of imaging quality, reduces the power supply requirements of the system, and enhances the system's noise resistance and resolution.
Smart Images
Figure BDA0005139657310000021 
Figure BDA0005139657310000031 
Figure BDA0005139657310000032
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of terahertz stack imaging, and in particular relates to a terahertz stack imaging system based on quantum cascade lasers and physically driven neural networks. Background Art
[0002] Phase imaging technology can quantitatively observe the internal structure or refractive index distribution of samples with weak absorption, and has developed rapidly in bands including visible light, X-rays, and terahertz. In terahertz biomedical imaging, phase information is also crucial for cancer diagnosis. Terahertz imaging technology is divided into two categories. One is based on terahertz time-domain spectroscopy. This technology is based on single-point scanning and light delay. The advantage is that the time-domain waveform of the terahertz wave can be directly obtained, and then the phase information of the sample can be obtained, but the imaging speed and resolution are limited; the other is based on terahertz continuous laser source, and usually an array detector is required to collect light intensity information, and then the phase information of the sample is restored from the light intensity information through reconstruction algorithms. Terahertz holographic imaging technology and terahertz stack imaging technology both belong to the latter. Among them, terahertz stack imaging is a branch of traditional spatial domain stack imaging. The characteristics of stack imaging are that no reference light is required, and large-area samples can be imaged at the same time, and the imaging area is not limited to the area of the detector.
[0003] The existing terahertz stack imaging technology has made progress in both experimental design and reconstruction algorithm. 12 Hz), all use far-infrared carbon dioxide gas lasers as light sources, usually use a beam of parallel light to pass through the aperture to generate an illumination beam, and use an array detector to collect light intensity data. At present, there are studies including continuous terahertz reflection stacking imaging methods, continuous terahertz wave double-object distance stacking imaging methods, continuous terahertz wave multi-object surface stacking phase contrast imaging methods, and high-resolution continuous terahertz wave stacking imaging methods that illuminate objects with divergent spherical waves. In the low-frequency terahertz band (<2THz), InP Gunn diodes and conical horn antennas are often coupled as light sources (frequency equal to 0.1THz), and single-point detectors are used to collect light intensity data. In terms of algorithms, the current terahertz stacking imaging algorithm is mainly based on the ePIE (extended Ptychographical Iterative Engine) algorithm. Based on the ePIE algorithm, current research includes terahertz stacking imaging probe position error correction methods and improving terahertz wave stacking imaging resolution through iterative extrapolation algorithms. At the same time, in other bands, traditional real-space domain stacking imaging algorithms also include DM (Difference Maps) algorithms. In addition to iterative algorithms, in the terahertz band, there is also a method to improve the image quality of terahertz stack imaging by using a deep learning neural network model based on deep blind degradation.
[0004] Experimentally, it includes a high-resolution continuous terahertz wave stack imaging method that uses a divergent spherical wave to illuminate an object. The divergent spherical wave is used to illuminate the object, effectively utilizing all the beam energy. However, due to the divergence of the beam, a larger array detector is required to collect data. The larger the area of the array detector in the terahertz band, the more expensive it will be. In the low-frequency terahertz band, a system of beam shaping elements is used to generate a terahertz beam for illumination. For example, the terahertz stack imaging method and system using a Bessel beam as a probe needs to be combined with a single-point detector for data collection, and the resolution and imaging speed of the system are limited.
[0005] The algorithm includes improving the resolution of terahertz wave stack imaging through an iterative extrapolation algorithm, which is essentially based on the traditional iterative algorithm of stack imaging, so its robustness is not strong enough; using a deep learning neural network model based on deep blind degradation to improve the image quality of terahertz stack imaging requires a large amount of training data to train the network, but due to the limitations of terahertz band devices, it is difficult to collect a large amount of data, so its training data can only be obtained through digital simulation methods.
[0006] Terahertz (THz) stack imaging usually uses a beam of parallel light to pass through an aperture to generate an illumination probe beam. However, since the aperture will filter out most of the terahertz waves, terahertz stack imaging is highly dependent on far-infrared gas lasers with high output power as light sources. Far-infrared gas lasers are large in size, expensive, and have high power supply requirements, which limits the practical application of terahertz stack imaging. At the same time, since the development of light source power and detectors in the terahertz band is not mature, this further limits the signal-to-noise ratio of the terahertz stack imaging system. Therefore, there is an urgent need to develop a terahertz stack imaging technology that can improve the utilization rate of terahertz waves in experiments and effectively improve the imaging quality in reconstruction algorithms. Summary of the invention
[0007] The technical problem to be solved by the present invention is to provide a terahertz stack imaging system based on quantum cascade lasers and physically driven neural networks, so as to improve the portability of the terahertz stack imaging system and the robustness of the reconstruction effect.
[0008] In order to solve the above technical problems, the technical solution of the present invention is:
[0009] A terahertz stack imaging system based on quantum cascade laser and physical driven neural network comprises a quantum cascade laser, a TPX lens, a two-dimensional translation stage, a terahertz array detector and a physical driven neural network structure; a sample is fixed on the two-dimensional translation stage, and the two-dimensional translation stage can move on an XY plane; the quantum cascade laser emits a terahertz wave, which passes through the TPX lens to form an illumination probe, and irradiates the sample, and the areas of the sample illuminated by the probe at adjacent translation positions overlap, the terahertz array detector collects light intensity information of the sample at each translation position, and the physical driven neural network structure reconstructs the amplitude and phase information of the sample using the light intensity information.
[0010] Preferably, the physical driven neural network structure includes a neural network module and a physical propagation module, and the collected N diffraction patterns I are input into the neural network module. Two Y-type neural networks are used to generate complex amplitude functions of the object O(r) and the probe P(r) respectively:
[0011] O(r)=|O(r)|·exp(jφ[O(r)])
[0012] P(r)=|P(r)|·exp(jφ[P(r)])
[0013] Among them, “||” and “φ[]” represent the spatial distribution calculation of amplitude and phase on the two-dimensional coordinate r = (x, y); the weights of the two Y-type neural networks are called W O and W P ;
[0014] The output of the neural network module is input to the physical propagation module, and the pupil constraint mask M is smoothed using a Gaussian filter and applied to the generated probe function; the diffraction pattern generated by the physical propagation module consists of a set of N diffraction patterns:
[0015] G n =|ASM(O(r―r n )·P(r)·M)| 2
[0016] Where M is the matrix used to constrain the area of the beam illuminating the sample, and the object O(r) is r n (n=1,2,...,N) translation, each diffraction pattern G generated n All are obtained by angular spectrum method;
[0017] When all physical parameters are defined, the resulting diffraction pattern is determined only by the weights of the two Y-neural networks:
[0018] G(W)={G 1 ,G 2,…,G N},W=(W O ,W P )
[0019] The reconstruction task of THz stylus imaging is formulated as solving the following problem:
[0020]
[0021] In order to obtain the best weight W * Finally, the complex amplitude functions of the object and the probe are reconstructed through the output of the physically driven neural network structure.
[0022] Preferably, the input of the neural network module is perturbed by adding Gaussian noise at each iteration:
[0023] I={I 1 ,I 2 ,…,I N},N=N1·N2
[0024] Where N1 and N2 represent the number of times the sample is translated along the x-axis and y-axis, respectively.
[0025] Preferably, the overlap ratio of samples between adjacent illumination positions is defined as:
[0026]
[0027] Where s represents the distance between adjacent lighting areas, and R represents the radius of the lighting area.
[0028] Preferably, the angular spectrum method is expressed by a mathematical formula:
[0029] ψ z＝d =F ―1 [H·F(ψ z＝0 )]
[0030] Among them, ψ z＝d and ψ z＝0 are the scalar components of the electromagnetic field at positions (x, y, d) and (x, y, 0) in three-dimensional space, respectively, and F represents the distance from the spatial coordinate (x, y) to the spatial frequency coordinate (f x ,f y ) is a two-dimensional Fourier transform mapping; the calculation formula of the original angular spectrum method transfer function H is:
[0031]
[0032] Where λ represents the wavelength and z represents the propagation distance along the optical axis; the transfer function under Fresnel diffraction conditions is approximately:
[0033]
[0034] Preferably, the mean square error between the collected diffraction pattern and the diffraction pattern generated by the physical propagation module is introduced as a loss function.
[0035] Compared with the prior art, the present invention has the following beneficial effects:
[0036] 1. The present invention uses a compact quantum cascade laser as a laser source and replaces the aperture in the traditional terahertz stack imaging with a lens, thereby obtaining a portable terahertz stack imaging system with low power supply requirements and high light utilization rate;
[0037] 2. The physics-driven deep learning algorithm adopted in the present invention does not require pre-training. It can not only improve the reconstruction quality, and has the advantages of strong anti-noise ability, high resolution, and good low overlap effect, but also can be easily migrated to real-space domain stacking imaging technology in other bands. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 It is a schematic diagram of a terahertz stack imaging system based on a quantum cascade laser and a physically driven neural network according to the present invention;
[0039] Figure 2 A schematic diagram of a terahertz stack imaging algorithm based on a physical driven neural network according to the present invention;
[0040] Figure 3 It is a specific parameter diagram of the Y-type neural network in the present invention;
[0041] Figure 4 Comparison of the imaging performance of UPNN, ePIE, and DM on three different experimental samples (binary, grayscale, and biomedical objects, respectively);
[0042] Figure 5 Comparison of the intensity distribution of the terahertz beam measured experimentally and reconstructed by the UPNN, ePIE and DM algorithms;
[0043] Figure 6 Experimental results of Siemens star resolution: ad) are the sample amplitudes reconstructed by the standard reference (GT) and three algorithms UPNN, ePIE and DM; e) represents the distribution of pixel values on the colored circles in (ad); fi) are the sample phases reconstructed by the standard reference (GT) and three algorithms UPNN, ePIE and DM; j) represents the distribution of pixel values on the colored circles in (fi);
[0044] Figure 7 Comparison of simulation results of terahertz-band stack imaging reconstructed by the standard reference GT and three algorithms: UPNN, ePIE and DM;
[0045] Figure 8 Comparison of simulation results of visible light band stack imaging reconstructed by standard reference GT and three algorithms: UPNN, ePIE and DM;
[0046] Fig. 9 Comparison chart of simulation reconstruction results under different noise and overlap conditions. DETAILED DESCRIPTION
[0047] The specific embodiments of the present invention are further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0048] like Figure 1 As shown, a terahertz stack imaging system based on quantum cascade laser and physical driven neural network includes quantum cascade laser 1, TPX lens 2, two-dimensional translation stage, terahertz array detector 4 and physical driven neural network structure (not shown in the figure); sample 3 is fixed on the two-dimensional translation stage, and the two-dimensional translation stage can move on the XY plane; quantum cascade laser 1 emits terahertz wave, which passes through the TPX lens with a focal length of 25cm to form an illumination probe, that is, a converging light beam 5, which is irradiated onto the sample 3, and the areas of the sample 3 irradiated by the probe at adjacent translation positions overlap, the terahertz array detector 4 collects the light intensity information of the sample 3 at each translation position, and the physical driven neural network structure reconstructs the amplitude and phase information of the sample using the light intensity information.
[0049] like Figure 2 As shown in Figure 1, the untrained physics-driven neural network structure (UPNN) consists of a neural network module and a physical propagation module. The neural network module takes a set of N collected diffraction patterns I as input. At each iteration, the input is perturbed by Gaussian noise to obtain a high-quality output image:
[0050] I={I 1 ,I 2 ,…,I N},N=N1·N2
[0051] Where N1 and N2 represent the number of times the sample is translated along the x-axis and y-axis, respectively. The overlap ratio between adjacent illumination positions is defined as:
[0052]
[0053] Where s represents the distance between adjacent lighting areas. Each lighting area is approximately a circle, that is, the distance between adjacent circle centers. R represents the radius of the lighting area.
[0054] The diffraction pattern input to the network can be obtained either through experimental measurement or numerical simulation. Figure 2 The simulated diffraction pattern is shown as input. Two Y-type neural networks are used to generate the complex amplitude functions of the object O(r) and the probe P(r), respectively:
[0055] O(r)=|O(r)|·exp(jφ[O(r)])
[0056] P(r)=|P(r)|·exp(jφ[P(r)])
[0057] Among them, “||” and “φ[]” represent the spatial distribution calculation of amplitude and phase on the two-dimensional coordinate r=(x,y). The weights of the two Y-type neural networks are called W O and W P The Adam optimizer with a learning rate of 0.001 is used to update the weights. The specific required parameters of the Y-type neural network are as follows Figure 3 shown.
[0058] The output of the neural network module is then fed into the physical propagation module, which is designed based on the data acquisition process of terahertz stack imaging. During the reconstruction of the probe function, the pixel values outside the circular probe function should always be zero. Therefore, a pupil constraint mask M with a radius of 3.25 mm (roughly equal to the experimentally measured probe diameter) is smoothed using a Gaussian filter and applied to the generated probe function. The diffraction pattern generated by the physical propagation module consists of a set of N diffraction patterns:
[0059] G n =|ASM(O(r―r n )·P(r)·M)| 2
[0060] Where M represents the matrix used to constrain the area of the light beam illuminating the sample, that is, the values of the matrix are 1 for the area whose distance from the center coordinate is less than the beam radius, and the values of the area outside the beam radius are 0; the object O(r) is r n (n=1,2,...,N) translation, each diffraction pattern Gn generated is obtained by Angular Spectrum Method (ASM). Angular Spectrum Method can be expressed by mathematical formula as follows:
[0061] ψ z＝d =F ―1[H·F(ψ z＝0 )]
[0062] Among them, ψ z＝d and ψ z＝0 are the scalar components of the electromagnetic field at positions (x, y, d) and (x, y, 0) in three-dimensional space, respectively, and F represents the distance from the spatial coordinate (x, y) to the spatial frequency coordinate (f x ,f y ). The calculation formula of the original angular spectrum transfer function H is:
[0063]
[0064] Where λ represents the wavelength and z represents the propagation distance along the optical axis. The transfer function under Fresnel diffraction conditions can be approximated as:
[0065]
[0066] The original angular spectrum transfer function and Fresnel transfer function are applied to the visible light and terahertz ranges, respectively.
[0067] Once all physical parameters are defined, the resulting diffraction pattern is completely determined by the weights of two Y-neural networks:
[0068] G(W)={G 1 ,G 2 ,…,G N},W=(W O ,W P )
[0069] The mean square error (MSE) between the collected diffraction patterns and the generated diffraction patterns is introduced as a loss function to optimize the weights of the neural network modules. For the proposed UPNN, the reconstruction task of terahertz stack imaging is formulated as solving the following problem:
[0070]
[0071] In order to obtain the best weight W * Finally, the complex amplitude functions of the object and the probe are reconstructed through the output of the UPNN neural network module.
[0072] The experimental results of terahertz stack imaging technology based on quantum cascade lasers and physically driven neural networks are as follows: Figure 4 As shown in Figure 1, the imaging effect of UPNN is better than that of the traditional ePIE and DM iterative algorithms. Using the angular spectrum method on the reconstructed probe, the spatial distribution of the illumination beam can be calculated, as shown in Figure 1. Figure 5As shown in Figure 2, it shows that UPNN has better reconstruction effect on the probe than the traditional ePIE and DM iterative algorithms. Figure 6 As shown in Figure 2. Through simulation, we can further quantitatively analyze that the imaging effect of UPNN is better than that of the traditional ePIE and DM iterative algorithms, as shown in Figure 2. Figure 7 As shown. Moreover, UPNN is also applied to the visible light band, such as Figure 8 UPNN can achieve high-quality reconstruction under high noise and low overlap conditions, such as Fig. 9 shown.
[0073] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.
Claims
1. A terahertz stack imaging system based on quantum cascade laser and physically driven neural network, characterized in that: It includes a quantum cascade laser, a TPX lens, a two-dimensional translation stage, a terahertz array detector and a physical driven neural network structure; the sample is fixed on the two-dimensional translation stage, and the two-dimensional translation stage can move on the XY plane; the quantum cascade laser emits a terahertz wave, which passes through the TPX lens to form an illumination probe and illuminates the sample, and the areas of the sample illuminated by the probe at adjacent translation positions overlap, the terahertz array detector collects the light intensity information of the sample at each translation position, and the physical driven neural network structure uses the light intensity information to reconstruct the amplitude and phase information of the sample.
2. The terahertz stack imaging system based on quantum cascade laser and physical driven neural network according to claim 1, characterized in that: The physical driven neural network structure includes a neural network module and a physical propagation module. The collected N diffraction patterns I are input into the neural network module. Two Y-type neural networks are used to generate the complex amplitude functions of the object O(r) and the probe P(r) respectively: O(r)=|O(r)|·exp(jφ[O(r)]) P(r)=|P(r)|·exp(jφ[P(r)]) Among them, "||" and "φ[]" represent the spatial distribution calculation of amplitude and phase on the two-dimensional coordinate r = (x, y); the weights of the two Y-type neural networks are called W O and W P ; The output of the neural network module is input to the physical propagation module, and the pupil constraint mask M is smoothed using a Gaussian filter and applied to the generated probe function; the diffraction pattern generated by the physical propagation module consists of a set of N diffraction patterns: G n =|ASM(O(r―r n )·P(r)·M)| 2 Where M is the matrix used to constrain the area of the beam illuminating the sample, and the object O(r) is r n (n=1,2,...,N) translation, each diffraction pattern G generated n All are obtained by angular spectrum method; When all physical parameters are defined, the resulting diffraction pattern is determined only by the weights of the two Y-neural networks: G(W)={G1,G2,…,G N },W=(W O ,W P ) The reconstruction task of THz stylus imaging is formulated as solving the following problem: In order to obtain the best weight W * Finally, the complex amplitude functions of the object and the probe are reconstructed through the output of the physically driven neural network structure.
3. The terahertz stack imaging system based on quantum cascade laser and physical driven neural network according to claim 2, characterized in that: At each iteration of the neural network module, the input is perturbed by adding Gaussian noise: I={I1,I2,…,I N },N=N1·N2 Where N1 and N2 represent the number of times the sample is translated along the x-axis and y-axis, respectively.
4. The terahertz stack imaging system based on quantum cascade laser and physical driven neural network according to claim 3, characterized in that: The overlap ratio of samples between adjacent illumination positions is defined as: Where s represents the distance between adjacent lighting areas, and R represents the radius of the lighting area.
5. The terahertz stack imaging system based on quantum cascade laser and physical driven neural network according to claim 2, characterized in that: The angular spectrum method is expressed mathematically as: ψ z＝d =F ―1 [H·F(ψ z＝0 )] Among them, ψ z＝d and ψ z＝0 are the scalar components of the electromagnetic field at positions (x, y, d) and (x, y, 0) in three-dimensional space, respectively, and F represents the distance from the spatial coordinate (x, y) to the spatial frequency coordinate (f x ,f y ) is a two-dimensional Fourier transform mapping; the calculation formula of the original angular spectrum method transfer function H is: Where λ represents the wavelength and z represents the propagation distance along the optical axis; the transfer function under Fresnel diffraction conditions is approximately:
6. The terahertz stack imaging system based on quantum cascade laser and physical driven neural network according to claim 2, characterized in that: The mean square error between the collected diffraction pattern and the diffraction pattern generated by the physical propagation module is introduced as a loss function.