Image processing metasurface design method based on complex amplitude modulation

Abstract

The invention belongs to the technical field of image processing, and provides an image processing metasurface design method based on complex amplitude modulation, which comprises a dielectric metasurface consisting of a silicon dioxide substrate and alpha-S nano-columns on the substrate, and adopts a dielectric nano-column arrangement method based on a geometric phase physical concept. Enabling the point spread function of the metasurface to be equal to a specific convolution kernel, and carrying out image processing operation according to the principle that the response of an image through a system corresponds to the convolution of the image and the point spread function of the system; according to the method, a dielectric nano-column arrangement method based on a geometric phase physical concept is adopted, so that the point spread function of the metasurface is equal to a specific convolution kernel, edge detection and second-order differential image processing operation are carried out by utilizing the principle that the response of an image through a system corresponds to the convolution of the image and the point spread function of the system, and the detection precision of the metasurface is improved. According to the method, the size of the optical convolution layer is greatly reduced, and the integration level and the operation speed are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular is an image processing super surface design method based on complex amplitude modulation. Background Art

[0002] At present, the in-depth research on deep convolutional neural networks (CNN) has enabled artificial intelligence (AI) to develop rapidly in academia and industry. In each layer of CNN, many kernels with specific functions are used for convolution operations to extract important features of objects for recognition, but as AI tasks increase, this method is still very time-consuming. Although advanced electronic devices such as graphics processing units, field programmable gate arrays, and tensor processing units have been developed to accelerate calculations to address these challenges, due to the limitations of electrical devices, computing speed and energy consumption are still affected by the low response of electronics, such as the charging and discharging of capacitors, electromagnetic radiation, and the movement of electrons in materials to generate heat.

[0003] Optical neural networks are considered to be the key to breaking through the bottleneck of the next generation of artificial neural networks. In theory, photons, as massless bosons, can propagate losslessly and at high speed through large bandgap transparent materials, and can quantitatively regulate optical information through optical diffraction structures, without the need for electronic processing modules such as analog-to-digital and digital-to-analog converters, thereby achieving near-light-speed and low-loss calculations. At present, optical neural networks can be divided into two categories: optoelectronic hybrid neural networks and all-optical neural networks. Among them, optoelectronic hybrid neural networks can realize the functions of convolutional neural networks, but the optical part of the network can only realize the function of convolution, and the electrical signal obtained after optoelectronic conversion continues to propagate in the electronic neural network. Although all-optical neural networks do not require the process of optoelectronic conversion, they cannot achieve the effect of convolution and can only complete the function of the fully connected layer. According to the different main optical components used, all-optical neural networks can be divided into three categories: photonic chips, passive diffraction optical elements, and scattering materials.

[0004] However, the optoelectronic hybrid neural network based on diffractive optical elements is similar to the principle of spatial filtering using a "4f system" composed of two convex lenses with focal lengths of f. A phase plate is placed in the Fourier plane in the middle of the "4f system" to modulate the amplitude and phase of the incident light to achieve convolution. The result of the optical operation still has a large amount of information. The all-optical nanophotonic circuit neural network is mainly based on the Mach-Zehnder interferometer with waveguide mode and phase modulation. The mathematical process is clearer than all other all-optical neural networks, but more devices are required to participate in the operation, the module size is larger, and it is not conducive to integration.

[0005] To this end, those skilled in the art have proposed an image processing metasurface design method based on complex amplitude modulation. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides an image processing metasurface design method based on complex amplitude modulation to solve the problems mentioned in the background technology.

[0007] A method for designing an image processing metasurface based on complex amplitude modulation includes a class of dielectric metasurfaces, which are composed of a silicon dioxide substrate and α-Si nanopillars on the substrate. A dielectric nanopillar arrangement method based on the concept of geometric phase physics is adopted to make the point spread function of the metasurface equal to a specific convolution kernel. The image processing operation is performed by using the principle that the response of an image passing through a system corresponds to the convolution of the image and the point spread function of the system.

[0008] By regulating the geometric phase and amplitude of the unit structure, the point spread function of the metasurface is made equivalent to the convolution kernel, and the image is convolved and processed;

[0009] The complex amplitude of vortex light with different orbital quantum numbers l and characteristic numbers p is used as the pupil function of the metasurface, and the pupil function of the metasurface, i.e. the complex amplitude of the metasurface, is modulated in real space. Since the size and azimuth angle of the nanocolumns can be arbitrarily controlled, the complex amplitude of the metasurface can be arbitrarily selected, and any image processing operation can be performed. Edge detection is used to convert the input Gaussian light beam into vortex light.

[0010] Preferably, under monochromatic light illumination, the distribution of a certain object plane image can be decomposed into countless delta functions, i.e., linear combinations of impulse functions, and each delta function can be used to obtain its corresponding point spread function. The coherent point spread function of the imaging system is derived by introducing the pupil function P(x, y). The distances between the metasurface and the object plane and the image plane are l1 and l2, respectively. The position of a point source on the object plane is set to (x0, y0), and the electric field distribution of the spherical wave emitted by it passing through the position l1 away from the point source is:

[0011]

[0012] Assuming that the complex pupil function of the metasurface is P(x1,y1), the distribution of the electric field after passing through the metasurface is expressed as:

[0013] u2(x1,y1)=u1(x1,y1)·P(x1,y1)

[0014] After Fresnel diffraction, the electric field distribution on the image plane at a distance of l2 from the metasurface is:

[0015]

[0016]

[0017] Preferably, an additional lens phase function is applied to the metasurface:

[0018]

[0019] Among them, f represents the focal length of the lens. According to the relationship between the focal length of the lens and the object distance and image distance: So that:

[0020]

[0021] make Then the integral term is equal to:

[0022] Where Φ(·) is the Fourier transform, then the electric field distribution u2′(x2,y2) on the image plane, i.e. the point spread function, is:

[0023]

[0024] After obtaining the point spread function, the image plane image is equal to the convolution of the original image and the point spread function, then the image plane field distribution is U(x2,y2)=U(x1,y1)*PSF∝U(x1,y1)*Φ[P(x1,y1)], where * is the convolution operation symbol.

[0025] Preferably, the Fourier transform form of the complex pupil function P(x, y) of the metasurface is made equivalent to a convolution kernel of a specific function, and the metasurface performs image processing and convolution operations. The metasurface performs filtering in the spatial frequency domain by regulating the complex amplitude of the real space of the metasurface.

[0026] Preferably, in response to circularly polarized light of 780 nm wavelength, the conversion efficiency of cross circular polarization and the geometric phase in the cross polarization term are controlled by regulating the length and width of the nanorods made of α-Si and with a period of 400 nm and the angle of the long axis relative to the x-axis, so as to achieve simultaneous regulation of amplitude and phase;

[0027] The circularly polarized light carrying the image information is transmitted to the metasurface at point l1 in front of the metasurface, and after being modulated by the metasurface, it finally forms a processed image at point l2 behind the metasurface. The cross-polarization component is detected by a half-wave plate and an analyzer to obtain the image information in the cross-circular polarization state.

[0028] Preferably, the edge detection image processing operation: using the electric field of the Laguerre-Gaussian beam as a reference for the complex pupil function of the metasurface to perform complex amplitude control, the amplitude of the donut shape and the vortex phase satisfy the conditions for performing a two-dimensional differential operation, thereby extracting the edges in the x and y directions;

[0029] Gaussian filtering image processing operation: The metasurface complex pupil function is modulated in the form of a Gaussian function, so that the metasurface performs a function similar to a low-pass filter, filtering out noise dominated by high-frequency signals.

[0030] Preferably, it also includes using elliptical nanocolumns made of amorphous silicon as the supersurface backup unit structure, based on the principle of Fano resonance, using the phase and amplitude mutations brought by the resonance to accurately control the amplitude and phase of the light field.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. The present invention can make the point spread function of the metasurface equivalent to the convolution kernel by regulating the geometric phase and amplitude of the unit structure, thereby realizing the convolution and processing functions of the image, with a computing speed close to the speed of light, small loss, and no large amount of heat loss; the integration is higher, and the manufacturing size is limited to the micron level; the unit structure of the metasurface is in a sub-wavelength size, which can provide an ultra-high image edge extraction accuracy of about 2.6μm and an energy utilization rate of about 0.7, and can achieve more efficient edge detection.

[0033] 2. The present invention modulates the pupil function of the metasurface, i.e., the complex amplitude of the metasurface, in real space. Since the size and azimuth of the nanocolumns can be arbitrarily controlled, the complex amplitude of the metasurface can also be arbitrarily selected, thereby realizing any image processing operation and satisfying the convolutional neural network's requirements for a variety of convolution kernels. At the same time, by utilizing the effect of edge detection, the input Gaussian light beam can be converted into vortex light, i.e., a donut-shaped intensity distribution and a vortex-shaped phase distribution, thereby realizing other functions besides image processing operations such as edge detection.

[0034] 3. The present invention uses vortex light complex amplitudes with different orbital quantum numbers l and characteristic numbers p as the pupil function of the metasurface to perform targeted information filtering on images containing different high-frequency information, so that the image only contains useful information. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 The working principle diagram of the convolutional hypersurface image processing of the present invention, (a) schematic diagram of PSF calculation of the convolutional hypersurface, where u1(x1, y1) and u2(x1, y1) represent the light field before and after the hypersurface, (b), (c) schematic diagram of the hypersurface used for detecting the edge of the input image (b) and Gaussian denoising (c);

[0036] Figure 2 The schematic diagram of the unit structure used for complex amplitude modulation of the present invention;

[0037] Figure 3Figure 1 is a diagram of the edge detection metasurface of the present invention. (a) and (c) are the amplitude and phase distribution of the metasurface under circularly polarized light illumination, respectively. The inset in (a) is the amplitude distribution after discretization. (b) is the SEM image of the prepared metasurface. The diameter of the entire metasurface is 600 μm and the scale bar is 2 μm. (d) is the measured intensity distribution of the metasurface. The insets show the simulated (curve) and measured (circle) intensity curves along the blue and orange dashed lines in (a) and (d), respectively.

[0038] Figure 4 Figure 1 is a Gaussian filtered metasurface image of the present invention. (a) and (c) are the amplitude and phase distribution of the metasurface under circularly polarized light illumination, respectively. The inset in (a) is the amplitude distribution after discretization. (b) is the SEM image of the prepared metasurface. The diameter of the entire metasurface is 600 μm and the scale is 2 μm. (d) is the measured intensity distribution of the metasurface. The insets show the simulated (curve) and measured (circle) intensity curves along the blue and orange dashed lines in (a) and (d), respectively.

[0039] Figure 5 The image and tangent intensity distribution diagram of the metasurface after the interaction with the letter NJU obtained by MATLAB numerical simulation in the experimental example of the present invention;

[0040] Figure 6 The edge detection results and tangent intensity distribution diagram of the letter NJU obtained by simulation and experiment of the present invention;

[0041] Figure 7 The figures are numerical simulation and experimental measurement results under different noises or patterns of the present invention, and the scale bar is 20 μm. DETAILED DESCRIPTION

[0042] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0043] Embodiment 1: The present invention provides an image processing metasurface design method based on complex amplitude modulation, including a dielectric metasurface composed of a silicon dioxide substrate and α-Si nanopillars on the substrate, and adopting a dielectric nanopillar arrangement method based on the concept of geometric phase physics to make the point spread function of the metasurface equal to a specific convolution kernel, and performing image processing operations by using the principle that the response of an image passing through a system corresponds to the convolution of the image and the point spread function of the system;

[0044] By regulating the geometric phase and amplitude of the unit structure, the point spread function of the metasurface is made equivalent to the convolution kernel, and the image is convolved and processed;

[0045] The complex amplitude of vortex light with different orbital quantum numbers l and characteristic numbers p is used as the pupil function of the metasurface, and the pupil function of the metasurface, i.e. the complex amplitude of the metasurface, is modulated in real space. Since the size and azimuth angle of the nanocolumns can be arbitrarily controlled, the complex amplitude of the metasurface can be arbitrarily selected, and any image processing operation can be performed. Edge detection is used to convert the input Gaussian light beam into vortex light.

[0046] like Figure 1 As shown in the figure, under monochromatic light illumination, the distribution of a certain object plane image can be decomposed into countless delta functions, that is, linear combinations of impulse functions, and each delta function can be used to calculate its corresponding point spread function. The point spread function is the image plane light field distribution corresponding to a point light source on the object plane after passing through the system. Taking a traditional lens as an example, P(x, y) is 0 at a position greater than the aperture, and the function value is 1 at a position less than the aperture. The pupil function can be a complex-valued function of position that varies to indicate the absorption of the system, thereby controlling the changes in the transmittance, reflectivity, phase, polarization, etc. of the incident light. For the metasurface, its complex amplitude distribution is the pupil function. By introducing the pupil function P(x, y), the coherent point spread function of the imaging system is derived. The distances between the metasurface and the object plane and the image plane are l1 and l2, respectively. The position of a point source on the object plane is set to (x0, y0), and the electric field distribution of the spherical wave emitted by it passing through the position l1 away from the point source is:

[0047]

[0048] Assuming that the complex pupil function of the metasurface is P(x1,y1), the distribution of the electric field after passing through the metasurface is expressed as:

[0049] u2(x1,y1)=u1(x1,y1)·P(x1,y1)

[0050] After Fresnel diffraction, the electric field distribution on the image plane at a distance of l2 from the metasurface is:

[0051]

[0052] Apply an additional lens phase function to the metasurface:

[0053]

[0054] Among them, f represents the focal length of the lens. According to the relationship between the focal length of the lens and the object distance and image distance: So that:

[0055]

[0056]

[0057] make Then the integral term is equal to:

[0058] Where Φ(·) is the Fourier transform, then the electric field distribution u2′(x2,y2) on the image plane, i.e. the point spread function, is:

[0059]

[0060] After obtaining the point spread function, the image plane image is equal to the convolution of the original image and the point spread function, then the image plane field distribution is U(x2,y2)=U(x1,y1)*PSF∝U(x1,y1)*Φ[P(x1,y1)], where * is the convolution operation symbol.

[0061] The Fourier transform form of the complex pupil function P(x, y) of the metasurface is made equivalent to a convolution kernel with a specific function, and the metasurface performs image processing and convolution operations. The metasurface performs filtering in the spatial frequency domain by regulating the complex amplitude of the real space of the metasurface, thereby removing unnecessary spatial frequencies and replacing the linear operation layer in the traditional convolutional neural network.

[0062] like Figure 2 As shown, in order to realize the specific complex amplitude function of the metasurface, the structure responds to circularly polarized light with a wavelength of 780nm. By adjusting the length and width of the nanorods made of α-Si and with a period of 400nm and the angle of the long axis relative to the x-axis, the cross circular polarization, that is, the conversion efficiency from left circle to right circle or right circle to left circle and the geometric phase in the cross polarization term are controlled, so as to achieve simultaneous regulation of amplitude and phase. Figure 2 The period of the unit structure is 400nm, its length and width are marked by L and W respectively, the height is fixed at 500nm, and θ is the orientation angle relative to the x-axis. The (L, W) and orientation angles on the xy plane can independently control the amplitude and phase of the transmitted light.

[0063] The circularly polarized light carrying the image information is transmitted to the metasurface at l1 in front of the metasurface, and after being modulated by the metasurface, it finally forms a processed image at l2 behind the metasurface. l1 and l2 satisfy the relationship The cross-polarization components are detected by a half-wave plate and a polarizer to obtain image information of the cross-circular polarization state.

[0064] like Figure 1 As shown in (b) and (c), the edge detection image processing operation is: the electric field of the Laguerre-Gaussian beam is used as the reference of the complex pupil function of the metasurface to perform complex amplitude control. The donut-shaped amplitude and vortex-shaped phase meet the conditions for two-dimensional differential operations, thereby extracting the edges in the x and y directions;

[0065] Gaussian filtering image processing operation: The metasurface complex pupil function is modulated in the form of a Gaussian function, so that the metasurface performs a function similar to a low-pass filter, filtering out noise dominated by high-frequency signals.

[0066] The metasurfaces corresponding to the above two examples have been prepared by micro-nano processing and experimentally characterized; the complex amplitude modulation and other information used for edge detection and Gaussian filtering are shown in Figure 2. Figure 3 and Figure 4 shown.

[0067] Figure 3 The edge detection metasurface; (a) and (c) are the amplitude and phase distribution of the metasurface under circularly polarized light illumination, respectively. The inset in (a) is the discretized amplitude distribution, (b) the SEM image of the prepared metasurface, the diameter of the entire metasurface is 600μm, the scale bar is 2μm, (d) the measured metasurface intensity distribution, the insets show the simulated (curve) and measured (circle) intensity curves along the blue and orange dashed lines in (a) and (d), respectively;

[0068] Figure 4 is a Gaussian filtering metasurface; (a) and (c) are the amplitude and phase distribution of the metasurface under circularly polarized light, respectively. The inset in (a) is the discretized amplitude distribution, (b) is the SEM image of the prepared metasurface, the diameter of the entire metasurface is 600μm, and the scale bar is 2μm, (d) is the measured metasurface intensity distribution. The insets show the simulated (curve) and measured (circle) intensity curves along the blue and orange dashed lines in (a) and (d), respectively;

[0069] Embodiment 2: This embodiment is basically the same as the previous embodiment, except that it also includes using elliptical nanocolumns made of amorphous silicon as the supersurface backup unit structure. Based on the principle of Fano resonance, the phase and amplitude mutations brought by the resonance are used to accurately control the amplitude and phase of the light field.

[0070] Experimental example: Compared with existing electrical neural networks, the speed of light conduction calculations is fast, and light loss is smaller than electrical loss. Compared with optical convolutional neural networks, the amount of optical calculations and the number of modules required to be integrated in the system are reduced. By using a spatial light modulator to apply circularly polarized light carrying image information to a micron-sized metasurface device, the amplitude and phase of the optical image can be modulated, thereby realizing the function of convolution with a higher degree of integration.

[0071] At present, the edge detection accuracy of most optical convolutional neural networks is about 2-7μm. The edge intensity half-maximum full width obtained by numerical simulation through MATLAB is defined as the detection accuracy, which can reach 1.1μm, and the normalized intensity of the edge can reach 0.7 and above. Figure 5The edge detection accuracy obtained by simulation and experimental measurement is 2μm and 2.6μm respectively. Figure 6 Compared with the prior art, a more accurate edge detection result can be obtained.

[0072] At the same time, in order to verify whether the present invention can be used for Gaussian filtering image processing operations, four images with different patterns or noises were selected in the experiment to evaluate the performance of the Gaussian filtering metasurface. Figure 7 The first and third columns from left to right are the input images of the simulation and experiment, respectively. In these four images, the letters "NJU", the handwritten digits "7" and "5" are added with salt and pepper noise, and Gaussian noise is added to the handwritten digit "7" to verify that the design is effective for a variety of noises. Figure 7 The second and fourth columns from left to right are the denoised image results in simulation and experiment, respectively. Obviously, most of the noise is removed after passing through the metasurface. For quantitative analysis, the present invention introduces the peak signal-to-noise ratio (PSNR) as a standard for measuring image quality; the field of view of the present invention is about one-sixth of the diameter of the entire device, which is approximately equal to 100μm, while the diameter of the entire device is 600μm and the focal length is 2.5mm, which can be captured by a smaller CMOS sensor and converted into a digital signal, which is more conducive to integration.

[0073] From the above, we can see that by purposefully selecting the size and rotation angle of the nanocolumns, specific geometric phase and amplitude modulation can be applied to different positions on the incident circularly polarized light field. In principle, the point spread function of the metasurface can correspond to any commonly used neural network convolution kernel. Based on this principle, metasurfaces for arbitrary image operations can be designed. Therefore, they are very compatible with convolutional neural networks and can replace the linear operation layer in convolutional neural networks.

[0074] Additionally, in order to provide a concise description of exemplary embodiments, all features of an actual embodiment (ie, those features that are not relevant to the best mode presently contemplated for carrying out the invention or those that are not relevant to implementing the invention) may not be described.

[0075] It will be appreciated that in the development of any actual implementation, as in any engineering or design project, numerous implementation-specific decisions may be made. Such a development effort may be complex and time-consuming, but will be a routine task of design, fabrication, and production for those of ordinary skill having the benefit of this disclosure without undue experimentation.

[0076] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for designing an image processing metasurface based on complex amplitude modulation, characterized in that: A dielectric metasurface is provided, which is composed of a silicon dioxide substrate and α-Si nanorods on the substrate. A dielectric nanorod arrangement method based on the concept of geometric phase physics is adopted to make the point spread function of the metasurface equal to a specific convolution kernel. The image processing operation is performed by using the principle that the response of an image passing through a system corresponds to the convolution of the image and the point spread function of the system. By regulating the geometric phase and amplitude of the unit structure, the point spread function of the metasurface is made equivalent to the convolution kernel, and the image is convolved and processed; The complex amplitude of vortex light with different orbital quantum numbers l and characteristic numbers p is used as the pupil function of the metasurface, and the pupil function of the metasurface, i.e. the complex amplitude of the metasurface, is modulated in real space. Since the size and azimuth angle of the nanocolumns can be arbitrarily controlled, the complex amplitude of the metasurface can be arbitrarily selected, and any image processing operation can be performed. Edge detection is used to convert the input Gaussian light beam into vortex light.

2. The method for designing an image processing metasurface based on complex amplitude modulation according to claim 1, characterized in that: Under monochromatic light illumination, the distribution of a certain object plane image can be decomposed into countless delta functions, that is, linear combinations of impulse functions, and each delta function can be used to calculate its corresponding point spread function. By introducing the pupil function P(x, y), the coherent point spread function of the imaging system is derived. The distances between the metasurface and the object plane and the image plane are l1 and l2 respectively. The position of a point source on the object plane is set to (x0, y0), and the electric field distribution of the emitted spherical wave passing through the position l1 away from the point source is: Assuming that the complex pupil function of the metasurface is P(x1,y1), the distribution of the electric field after passing through the metasurface is expressed as: u2(x1,y1)=u1(x1,y1)·P(x1,y1) After Fresnel diffraction, the electric field distribution on the image plane at a distance of l2 from the metasurface is:

3. The method for designing an image processing metasurface based on complex amplitude modulation according to claim 2, characterized in that: Apply an additional lens phase function to the metasurface: Among them, f represents the focal length of the lens. According to the relationship between the focal length of the lens and the object distance and image distance: So that: make Then the integral term is equal to: Where Φ(·) is the Fourier transform, then the electric field distribution u2′(x2,y2) on the image plane, i.e. the point spread function, is: After obtaining the point spread function, the image plane image is equal to the convolution of the original image and the point spread function, then the image plane field distribution is U(x2,y2)=U(x1,y1)*PSF∝U(x1,y1)*Φ[P(x1,y1)], where * is the convolution operation symbol.

4. The method for designing an image processing metasurface based on complex amplitude modulation according to claim 3, characterized in that: The Fourier transform form of the complex pupil function P(x, y) of the metasurface is made equivalent to a convolution kernel of a specific function, and the metasurface performs image processing and convolution operations. The metasurface performs filtering in the spatial frequency domain by regulating the complex amplitude of the real space of the metasurface.

5. The method for designing an image processing metasurface based on complex amplitude modulation according to claim 4, characterized in that: Responding to circularly polarized light of 780nm wavelength, the conversion efficiency of cross circular polarization and the geometric phase in the cross polarization term are controlled by adjusting the length and width of the nanorods made of α-Si and with a period of 400nm, as well as the angle between the long axis and the x-axis, so as to achieve simultaneous regulation of amplitude and phase. The circularly polarized light carrying the image information is transmitted to the metasurface at l1 in front of the metasurface, and after being modulated by the metasurface, it finally forms a processed image at l2 behind the metasurface. l1 and l2 satisfy the relationship The cross-polarization components are detected by a half-wave plate and a polarizer to obtain image information of the cross-circular polarization state.

6. The method for designing an image processing metasurface based on complex amplitude modulation according to claim 5, characterized in that: Edge detection image processing operation: The electric field of the Laguerre-Gaussian beam is used as a reference for the complex pupil function of the metasurface to perform complex amplitude control. The donut-shaped amplitude and vortex-shaped phase meet the conditions for two-dimensional differential operations, thereby extracting edges in the x and y directions. Gaussian filtering image processing operation: The metasurface complex pupil function is modulated in the form of a Gaussian function, so that the metasurface performs a function similar to a low-pass filter, filtering out noise dominated by high-frequency signals.

7. The method for designing an image processing metasurface based on complex amplitude modulation according to claim 1, characterized in that: It also includes the use of elliptical nanocolumns made of amorphous silicon as the backup unit structure of the metasurface. Based on the principle of Fano resonance, the phase and amplitude mutations brought by resonance are used to precisely control the amplitude and phase of the light field.

Images