Deep Shift-Frequency Label-Free Super-Resolution Imaging Chip Based on Metasurface Film, and Its System and Method

Abstract

The present invention discloses a deep frequency-shifted label-free super-resolution imaging chip based on a metasurface thin film, as well as a system and method thereof, belonging to the field of frequency-shifted super-resolution microscopy imaging. The chip includes a substrate that is transparent in the illumination light band and has flat surfaces on both sides; the upper surface of the substrate has a number of microstructures for exciting surface plasmons with ultra-high wave vectors, and a metasurface thin film that can cover all the microstructures is also provided on the upper surface of the substrate; the metasurface thin film is composed of a longitudinal periodic arrangement of a metal film and a non-metal film, and is used to support the transmission of surface plasmons with ultra-high wave vectors. By controlling the types of the metal thin film and the non-metal thin film materials, the thickness of each layer of the thin film, the total thickness of the metasurface, and the illumination wavelength, etc., the present invention can design a hyperbolic metasurface with an effective refractive index much higher than that of natural optical materials, thereby further improving the resolution of the frequency-shifted super-resolution imaging technology and achieving sub-50-nanometer super-resolution imaging.

Description

Technical Field

[0001] The present invention belongs to the field of shifted-frequency super-resolution microscopy imaging, and particularly relates to a deep-shifted-frequency label-free super-resolution imaging chip based on a metasurface thin film, and a system and method thereof. Background Art

[0002] Imaging and detection with large field of view, high speed, and high resolution have important practical significance. The shifted-frequency super-resolution imaging technology can break through the bandwidth limitation of traditional imaging and detection devices and has significant advantages in fast large-field imaging. The shifted-frequency super-resolution imaging technology uses a natural waveguide with a high effective refractive index to transmit a high-wavevector evanescent field, modulates the high-frequency spatial information of the sample to a low-frequency region acceptable to the microscope, and then restores the high-frequency spatial information through a spatial spectrum iteration algorithm, and finally reconstructs a super-resolution image that breaks through the diffraction limit.

[0003] For example, in the Chinese patent application with the publication number CN110658195B and the invention name of a shifted-frequency label-free super-resolution microscopy chip and an imaging method thereof, based on a bulk high-refractive-index waveguide material, by preparing a grating structure on the waveguide surface, using the diffraction effect of the grating to couple free-space light into the waveguide material, and generating a transverse high-wavevector evanescent field by total internal reflection on the upper surface to illuminate the sample, so as to obtain the high-frequency information of the sample, and finally restore the super-resolution image of the sample. However, natural optical waveguide materials have limited refractive indices in the visible light band (n = 3.49 @ λ = 532 nm), and it is difficult to generate an evanescent field with a higher wavevector (k max = 3.49k0, k0 = 2π / λ, λ is the illumination wavelength) under linear conditions, thus limiting the improvement of the highest resolution of the shifted-frequency super-resolution imaging. Summary of the Invention

[0004] In order to solve the problem that natural optical waveguide materials have limited refractive indices in the visible light band, the present invention proposes a deep-shifted-frequency label-free super-resolution imaging chip based on a metasurface thin film, and a system and method thereof.

[0005] The metasurface thin film of the present invention is composed of a longitudinally periodic metal thin film and a non-metal thin film. The metal material can be Ag, Au, Cu, etc. with a negative refractive index in the visible light band, and the non-metal material can be SiO2, Al2O3, TiO2, etc. that are transparent in the visible light band. By controlling the types of the metal thin film and the non-metal thin film materials, the thickness of each layer of the thin film, the total thickness of the metasurface, and the illumination wavelength, etc., a hyperbolic metasurface with an effective refractive index much higher than that of natural optical materials can be designed, thereby further improving the resolution of the shifted-frequency super-resolution imaging technology and realizing sub-50-nanometer super-resolution imaging.

[0006] The specific technical solutions adopted by the present invention are as follows:

[0007] In a first aspect, the present invention provides a deep frequency-shifted label-free super-resolution imaging chip based on a metasurface thin film, comprising a substrate that is transparent in the illumination light band and has flat surfaces on both sides; the upper surface of the substrate has several microstructures for exciting surface plasmons with ultra-high wave vectors (e.g., k spp = 10k0), and a metasurface thin film capable of covering all the microstructures is further provided on the upper surface of the substrate; the metasurface thin film is composed of a longitudinal periodic arrangement of a metal film and a non-metal film and is used to support the transmission of surface plasmons with ultra-high wave vectors.

[0008] At the same time, through the thickness design of different film layers of the metasurface, surface plasmons with different wave vectors that can be transmitted can be selected according to the phase matching condition, and frequency-shifted information at different frequency-shift amounts can be obtained.

[0009] Preferably, when the illumination light is visible light, the substrate material can be silica (SiO2), alumina (Al2O3), or other waveguide materials that are transparent in the visible light band. The metal film material can be selected from one of gold (Au), silver (Ag), copper (Cu), etc. that have a negative refractive index in the visible light band, and the non-metal film material is one of silica (SiO2), magnesium fluoride (MgF2), and alumina (Al2O3).

[0010] The specific selection of the metal film and non-metal thin film materials, thickness, and total number of layers are determined by the illumination wavelength, imaging resolution requirements, processing conditions, etc. The thickness of each film layer is between a few nanometers and dozens of nanometers, and the minimum thickness is determined by the processing accuracy (e.g., the minimum thickness of the Ag film is 10 nanometers, and the minimum thickness of the SiO2 film is 4 nanometers). The specific thickness is determined by the required wave vector magnitude.

[0011] Preferably, the microstructures are arranged on the upper surface of the substrate by means of micro-nano processing, and the method is one of an electron beam lithography system EBL (E-Beam Lithography) electron beam direct writing system, focused ion beam (FIB) technology, lithography technology, nanoimprint technology, etc. The width, shape, etc. of the microstructures are determined according to specific requirements.

[0012] Preferably, the metasurface thin film is deposited on the upper surface of the substrate by means of micro-nano processing, and the methods include magnetron sputtering, electron beam evaporation and other technologies.

[0013] In a second aspect, the present invention provides an optical system, comprising an illumination light source that is incident on a chip to generate a lateral large-wavevector evanescent field, a deep-shift-frequency label-free super-resolution imaging chip based on a metasurface thin film as described in any one of the first aspects, a microscopic optical system for collecting scattered light of a sample, and an optical receiving system for recording original imaging information; the light source can be incident from the lower surface of the chip, illuminate the microstructures, excite large-wavevector surface plasmons, be transmitted through the metasurface thin film, and illuminate the sample above the metasurface thin film.

[0014] Preferably, a polarizer and a phase plate assembly can be added after the light source to change the polarization characteristics of the light source and improve the imaging signal-to-noise ratio.

[0015] Preferably, the light source is a laser light source or an LED light source.

[0016] Preferably, a beam shrinking and expanding system, a polarizer and a phase plate assembly, and a reflector are sequentially arranged on the optical path between the light source and the lower surface of the chip; the chip is placed on a sample stage, and the optical receiving system includes a lens, an optical camera, and a computer for storing, reconstructing, and displaying images.

[0017] In a third aspect, an imaging method using the optical system described in any one of the second aspects is as follows:

[0018] S11: Directly illuminate the sample with the light source, and use the optical camera in the optical receiving system to collect a fundamental frequency image including the low-frequency spatial information of the sample;

[0019] S12: Incident the illumination light generated by polarization modulation of the light source from the lower surface of the chip, excite surface plasmons with different wavevectors through the microstructures and transmit them in the metasurface thin film, generate a lateral large-wavevector evanescent wave above the metasurface thin film to illuminate the sample; change the polarization direction and wavelength of the illumination light, etc., and collect a shifted-frequency image containing the high-frequency spatial information of the sample in different directions and different frequencies;

[0020] S13: Input the collected fundamental frequency image and the shifted-frequency image into a trained deep learning restoration algorithm, and use the collected low-frequency spatial information and part of the high-frequency spatial information (shifted-frequency images under illumination by surface plasmons with different wavevectors) to restore the complete spectral information within the highest shifted-frequency range, and then reconstruct the deep-shift-frequency super-resolution image of the sample.

[0021] Preferably, deep learning adopts a deep Fourier domain attention neural network architecture (DFCAN), starting from the spectral domain, which is more universal than traditional deep learning algorithms in the spatial domain.

[0022] Preferably, the deep Fourier domain attention convolutional neural network architecture is first trained using shifted-frequency simulation data, and then further optimized using the shifted-frequency low-resolution images obtained under a low-magnification objective lens and the high-resolution images obtained under a high-magnification objective lens. Compared with the frequency-domain iterative algorithm, the deep learning neural network is faster, has a higher signal-to-noise ratio when processing a large amount of image data, and the number of shifted-frequency images required can be reduced to one-tenth of the original.

[0023] The deep Fourier domain attention convolutional neural network described in the present invention is composed of three parts: shallow low-frequency feature extraction, residual group high-frequency feature extraction, and an upsampling module. The low-frequency feature extraction module is implemented by a convolutional layer and a GELU activation function. The high-frequency feature extraction module is composed of m residual groups connected in series, and each residual group is composed of n Fourier channel attention blocks (FCAB). m and n are positive integers, which are determined by weighing the model performance and training efficiency. In addition, skip connections are introduced in each residual group to allow low-frequency features to pass directly through the network, enabling the FCAB to focus on the extraction of high-frequency feature information. Finally, the upsampling module upsamples the feature map to the same size as the ground truth image according to the magnification factor. This deep learning model uses the mean square error (MSE) and structural similarity (SSIM) as the joint loss function and uses an adaptive algorithm to minimize the loss function.

[0024] Furthermore, the Fourier channel attention mechanism used by this network includes a fast Fourier transform layer to calculate the power spectrum of the sample high-frequency features extracted by the residual group, and adaptively rescales and weights each feature map by integrating the contributions of all frequency components in the power spectrum, thereby allowing the network to use the power spectrum features of different feature maps to learn the feature mapping of sample high-frequency information. The present invention combines this deep Fourier domain attention convolutional neural network with the shifted-frequency theory, taking advantage of the network's ability to effectively learn the precise mapping of different features of the sample in the frequency domain and the shifted-frequency effect's ability to translate the high-frequency components in the sample frequency domain from multiple directions to the system passband range, enabling the deep learning model to accurately and effectively learn the high-frequency features of the sample.

[0025] Preferably, the training and reconstruction process of the deep learning specifically includes the following steps:

[0026] S21: Collect or generate high-resolution simulation images, which should have a high structural complexity to contain as many spatial features of the required reconstructed experimental samples as possible, such as points, straight lines, curves, and line widths, large-angle deflection structures, and channel structures.

[0027] S22: Define the physical parameters required for the deep shift-frequency super-resolution microscopy imaging model. The physical parameters include the numerical aperture of the objective lens, the illumination light wavelength, the shift frequency provided by the evanescent wave, the illumination angle, the spatial scale corresponding to a unit pixel, and the image magnification factor.

[0028] S23: According to the physical parameters defined in step S22, generate the corresponding coherent transfer function (CTF). Then, perform a fast Fourier transform on the high-resolution simulation image to obtain its spectrum, and simulate ordinary light and evanescent wave illumination to obtain a set of low-resolution image data.

[0029] That is, use the CTF to intercept the information of the sub-aperture at the corresponding position on the spectrum, perform an inverse Fourier transform on the intercepted sub-aperture to obtain the corresponding spatial domain information, and take its intensity as an illumination low-resolution image. Obtain multiple images with different shift frequencies according to the shift frequency provided by the evanescent wave and the illumination angle, and add the fundamental frequency image under ordinary light illumination to obtain a set of low-resolution image data.

[0030] S24: Use Gaussian white noise with different signal-to-noise ratios to process the low-resolution images in the dataset respectively to make them closer to the actual imaging conditions and enhance the denoising ability of the model. In addition, random cropping, horizontal / vertical flipping, and rotation transformations can be applied to further enrich the simulation dataset.

[0031] S25: Randomly divide the simulation dataset obtained after data augmentation in step S24 into a training set, a validation set, and a test set according to a certain ratio, and train the deep Fourier domain attention convolutional neural network.

[0032] S26: In the experimental environment, obtain a small amount of shift-frequency low-resolution images and corresponding high-resolution images under a low-magnification objective lens and a high-magnification objective lens as the training set, and on the basis of training with the simulation dataset, further fine-tune the deep learning model through transfer learning.

[0033] S27: Collect the fundamental frequency image and the shift-frequency image of the required reconstruction sample, and input them into the trained deep learning model for reconstruction.

[0034] The present invention has the following beneficial effects compared with the prior art:

[0035] The present invention solves the problem that the shift-frequency super-resolution imaging is limited by the limited refractive index of the natural optical waveguide, and uses the hyperbolic super surface (i.e., the metasurface thin film) to achieve a higher effective refractive index, and improves the shift-frequency super-resolution imaging resolution to sub-50 nanometer resolution. At the same time, the deep Fourier domain attention convolutional neural network algorithm is adopted, which reduces the original acquisition data and speeds up the reconstruction speed of the ultra-high resolution image, and is conducive to realizing fast and ultra-high resolution deep shift-frequency super-resolution microscopy imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the cross-section of a deep frequency-shifted label-free super-resolution imaging chip and the excitation light.

[0037] Figure 2 It is a schematic diagram of the dispersion characteristic curve and transmission efficiency of the Ag / SiO2 multi-layer metasurface thin film. The thickness of the Ag metal layer is taken as 13.7 nm, and the thickness of the SiO2 non-metal layer is 6 nm, where

[0038] Figure 3 It is a schematic diagram of the structure of the optical system: Among them, 201 is the light source, 202 is the beam reducing and expanding system, 203 is the polarizer and phase plate assembly, 204 is the mirror, 205 is the chip, 206 is the sample stage, 207 is the microscopic optical system, 208 is the optical camera, and 209 is the computer.

[0039] Figure 4 It is a schematic diagram of the comparison of the spectral detection range and resolution between a common microscope and a frequency-shifted imaging chip: (a) Comparison of the spectral detection range of a common microscope (numerical aperture is 0.8), a frequency-shifted imaging chip based on a natural waveguide chip (GaN, refractive index 2.55 @ 405 nm wavelength), and a deep frequency-shifted chip based on a metasurface thin film; (b-d) Comparison of the theoretical imaging capabilities of a common microscope, a frequency-shifted imaging chip based on a natural waveguide chip, and a deep frequency-shifted chip based on a metasurface thin film. Specific embodiments

[0040] The present invention will be further elaborated and described below in conjunction with the accompanying drawings and specific embodiments. The technical features of each embodiment in the present invention can be combined correspondingly without conflict.

[0041] As Figure 1 shown, a deep frequency-shifted label-free super-resolution imaging chip based on a metasurface thin film provided by the present invention includes: a substrate material with two flat surfaces that are transparent in the illumination light band. If the illumination light is visible light, the substrate material can be SiO2, Al2O3, or other waveguide materials that are transparent in the visible light band. The upper surface of the substrate material has periodic microstructures for exciting ultra-high wavevector surface plasmons, and the microstructures are processed by means such as FIB, EBL, lithography, and nanoimprinting; the upper surface of the substrate material is a metasurface thin film composed of a longitudinal periodic metal film and a non-metal film to support the transmission of ultra-high wavevector surface plasmons. If the illumination light is visible light, the metal material can be selected from Ag, Au, Cu, etc. that have a negative refractive index in the visible light band, and the non-metal material can be selected from SiO2, MgF2, Al2O3, etc. The specific material selection, thickness, and total number of layers of the metal thin film and non-metal thin film are determined by the illumination wavelength, imaging resolution requirements, processing conditions, etc.

[0042] In this embodiment, taking the structure in one-dimensional direction as an example, as Figure 1 shown. The illumination light is incident from the lower surface of the substrate material, and surface plasmons with different wave vectors are excited at the junction of the micro-structure and the metal layer. Subsequently, they are transmitted in the hyperbolic metamaterial surface thin film (i.e., the metamaterial surface thin film) and form transverse evanescent waves with different wave vectors at the upper surface of the metamaterial surface thin film to illuminate the sample. By changing the incident angle of the illumination light, the distribution of the illumination light field is changed, and by changing the polarization direction of the illumination light, the wave vector direction of the large wave vector transverse evanescent wave is regulated.

[0043] In the metamaterial surface thin film, the transmission dispersion characteristic equation of the ultra-large wave vector evanescent wave is

[0044]

[0045] In the formula, d m is the thickness of the metal layer, d d is the thickness of the non-metal layer, ε m is the dielectric constant of the metal layer, ε d is the dielectric constant of the non-metal layer, k zm is the longitudinal wave vector of the evanescent wave in the metal layer, k zd is the longitudinal wave vector of the evanescent wave in the non-metal layer, k B is the equivalent longitudinal wave vector of the evanescent wave in the longitudinally periodic metamaterial surface thin film. Combining with the equivalent refractive index model, the transmission efficiency in the periodic metamaterial thin film can be obtained as

[0046]

[0047] Among them, ω is the frequency of the input light, x is the transverse wave vector of the input light, k′ z is the longitudinal wave vector of the input light in the ambient medium, d is the thickness of the periodic metamaterial thin film, ε x is the equivalent transverse refractive index of the metamaterial surface, and ε is the refractive index of the ambient medium. The dielectric constants of metals such as Ag, Au, and Cu with negative refractive indices in the visible light band change violently with the wavelength. Therefore, the regulation of the maximum frequency shift amount can be achieved by regulating the wavelength.

[0048] Taking the multi-layer thin film composed of Ag and SiO2 as an example, if the thickness of the Ag metal layer is 12 nanometers and the thickness of the SiO2 non-metal layer is 6 nanometers. Assuming that the incident lights are common lasers of 405 nanometers, 532 nanometers, and 635 nanometers respectively, the dispersion curves and transmission efficiencies of the excited evanescent waves in the metamaterial surface are as Figure 2As shown. According to the phase matching condition, the metasurface supports the transmission of surface plasmons with multiple different wave vectors, and these surface plasmons with different wave vectors determine the spectral range of the frequency-shifted images that can be obtained by the system. At the same time, within the visible light range, the shorter the wavelength of the incident light, the larger the transverse wave vector of the evanescent wave that the metasurface thin film can support for transmission, and the higher the resolution.

[0049] The present invention also provides an optical system structure for imaging using a deep frequency-shifted label-free super-resolution imaging chip based on a metasurface thin film, as Figure 3 shown. This optical system structure includes: a light source 201, a beam shrinking and expanding system 202, a polarizer and phase plate assembly 203 for adjusting the polarization direction, a mirror 204; a sample stage 206 for supporting the chip, a microscopic optical system 207 for collecting the scattered light of the sample, an optical camera 208 for collecting the scattered light signal, and a computer 209 for storing, recording, storing, and processing the scattered light signal. The light source can be a laser or LED light.

[0050] The present invention also provides an imaging method for imaging using the above optical system structure, and this method includes:

[0051] Step 1: Illuminate the sample with a conventional microscope, and use an optical camera to collect the fundamental frequency image of the low-frequency spatial information of the sample;

[0052] Step 2: Adjust the polarizer and phase plate so that the illumination light is incident on the imaging chip with TM polarization, excite surface plasmons at the microstructures, and form a large wave vector transverse evanescent field through the metasurface to illuminate the sample above the metasurface. Collect the frequency-shifted images of the high-frequency spatial information of the sample in different directions through the optical receiving system;

[0053] Step 3: Adjust the wavelength of the illumination light, repeat the previous step, and utilize the dispersion characteristics of the metasurface thin film to collect the frequency-shifted images containing the high-frequency information of the sample at different frequencies through the optical receiving system.

[0054] Step 4: Input the fundamental frequency images and high-frequency images collected in Step 1, Step 2, and Step 3 into the trained deep frequency-shifted depth Fourier domain attention neural network, and utilize the collected low-frequency information and partial high-frequency information to restore the complete spectral information of the sample within the highest frequency-shifted range, and further reconstruct the deep frequency-shifted super-resolution image of the sample.

[0055] In this embodiment, SiO2 material (refractive index 1.4696 @ 405 nm wavelength) is used as the substrate material. A rectangular grating with a length and width of 500 microns is etched on the upper surface of the substrate. The grating period is 500 nm, the slit width is 50 nm, and the excitation light with a wavelength of 405 nm is used. Then, a total of 10 layers of metasurface films composed of a 13.7 nm Ag metal layer and a 6 nm SiO2 non-metal layer are deposited on the substrate. The equivalent refractive index can reach 12.1 at 405 nm. The NA of the objective lens is 0.8, and a resolution of 31.4 nm can be obtained, as Figure 4 shown, which is 16.1 times higher than the diffraction limit.

[0056] The above-described embodiments are only a preferred solution of the present invention, but they are not intended to limit the present invention. Those of ordinary skill in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all technical solutions obtained by means of equivalent replacement or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A deep frequency shift label-free super-resolution imaging chip based on a metasurface thin film, characterized in that A substrate that is transparent in the illumination light band and has flat surfaces on both sides; the upper surface of the substrate has several microstructures for exciting ultra-high wavevector surface plasmons, and a metasurface thin film covering all the microstructures is also provided on the upper surface of the substrate; the metasurface thin film is composed of longitudinally periodic arrangements of a metal film and a non-metal film, and is used to support the transmission of ultra-high wavevector surface plasmons. When the illumination light is visible light, the substrate material is silica or alumina, the metal film material is one of gold, silver, and copper, and the non-metal film material is one of silica, magnesium fluoride, and alumina.

2. The deep shift-frequency label-free super-resolution imaging chip based on the metasurface thin film according to claim 1, characterized in that The microstructures are arranged on the upper surface of the substrate by micro-nano processing methods, and the methods are one of an electron beam lithography system, a focused ion beam technology, a lithography technology, and a nanoimprint technology.

3. The deep frequency shift label-free super-resolution imaging chip based on the metasurface thin film according to claim 1, wherein The metasurface thin film is deposited on the upper surface of the substrate by micro-nano processing methods, and the methods are magnetron sputtering or electron beam evaporation.

4. An optical system, characterized in that, It includes a light source (201), a deep shift-frequency label-free super-resolution imaging chip (205) based on a metasurface thin film according to any one of claims 1 to 3, a microscopic optical system (207), and an optical receiving system; the light source (201) can be incident from the lower surface of the chip (205), illuminate the microstructures to excite large wavevector surface plasmons, transmit through the metasurface thin film, and illuminate the sample above the metasurface thin film; the microscopic optical system (207) is used to collect the scattering signals of the sample, and the optical receiving system is used to record the original imaging information.

5. The optical system according to claim 4, characterized in that, The light source (201) is a laser light source or an LED light source.

6. The optical system according to claim 4, wherein A beam shrinking and expanding system (202), a polarizer and a phase plate assembly (203), and a mirror (204) are successively provided on the optical path between the light source (201) and the lower surface of the chip (205); the chip (205) is placed on a sample stage (206), and the optical receiving system includes a lens, an optical camera (208), and a computer (209) for storing, reconstructing, and displaying images.

7. An imaging method using the optical system according to any one of claims 4 to 6, characterized in that, Specifically as follows: S11: Use the light source (201) to directly illuminate the sample, and use the optical camera (208) in the optical receiving system to collect a fundamental frequency image including the low-frequency spatial information of the sample. S12: Incident the illumination light generated by polarization modulation of the light source (201) from the lower surface of the chip (205), excite surface plasmons with different wavevectors through the microstructures and transmit in the metasurface thin film, and generate a transverse large wavevector evanescent wave above the metasurface thin film to illuminate the sample; change the polarization direction and wavelength of the illumination light, and collect a shifted frequency image containing the high-frequency spatial information of the sample in different directions and different frequencies. S13: Input the collected fundamental frequency image and the shifted frequency image into a trained deep learning restoration algorithm, use the collected low-frequency spatial information and part of the high-frequency spatial information to restore the complete spectral information in the highest shifted frequency range, and then reconstruct the deep shift-frequency super-resolution image of the sample.

8. The imaging method according to claim 7, wherein The deep learning restoration algorithm adopts a deep Fourier domain attention neural network architecture.

9. The imaging method according to claim 7, wherein The training and reconstruction processes of the deep learning specifically include the following steps: S21: Collect or generate high-resolution simulation images that contain the spatial features of the required reconstruction experimental samples; S22: Define the physical parameters required for the deep shift-frequency super-resolution microscopy imaging model, where the physical parameters include the numerical aperture of the objective lens, the illumination light wavelength, the shift-frequency amount provided by the evanescent wave, the illumination angle, the spatial scale corresponding to a unit pixel, and the image magnification factor; S23: According to the physical parameters defined in step S22, generate the corresponding coherent transfer function, then perform a fast Fourier transform on the high-resolution simulation image to obtain its spectrum, and simulate ordinary light and evanescent wave illumination to obtain a set of low-resolution image data; S24: Process the low-resolution image data in the dataset using Gaussian white noise with different signal-to-noise ratios respectively to make it closer to the actual imaging conditions and enhance the denoising ability of the model; at the same time, apply random cropping, horizontal or vertical flipping, and rotation transformations to further enrich the simulation dataset; S25: Randomly divide the simulation dataset obtained after data augmentation in step S24 into a training set, a validation set, and a test set according to a ratio, and train the deep Fourier domain attention convolutional neural network; S26: In the experimental environment, obtain shift-frequency low-resolution images and corresponding high-resolution images under a low-magnification objective lens and a high-magnification objective lens as the training set, and on the basis of training with the simulation dataset, further fine-tune the deep learning model through transfer learning; S27: Collect the fundamental frequency image and the shift-frequency image of the required reconstruction sample, and input them into the trained deep learning model for reconstruction.

Images