A metasurface device with polarization-switchable image recognition function and implementation method thereof

Abstract

The present invention relates to a metasurface device with polarization-switchable image recognition function and an implementation method thereof. The metasurface device comprises a metal digital imaging plate, a metasurface layer and a detection plane. The metasurface layer comprises a dielectric substrate, a silicon column structure array placed thereon and a metal grating structure on the back. A diffraction neural network is used to simulate the real optical diffraction process, modulate the incident light wave, and optimize and calculate the phase distribution of the diffraction layer using gradient descent and back propagation algorithms. According to the optimized diffraction layer phase parameters, dynamics and geometric phase principles are applied, eight discrete rectangular silicon columns are selected and rotated 45 degrees to generate a metasurface structure, and a metal grating structure is applied on the back side to realize the polarization-switchable image classification function based on a neural network, which solves the problem of the single function of the current metasurface device, and can identify and classify according to the incident linear polarized light of different polarization directions on both sides of the metasurface device, significantly improving the efficiency in the detection task.

Description

Technical Field

[0001] The present invention relates to a metasurface device, belonging to the fields of optical neural network, micro-nano optics, optical information processing and machine vision, and in particular to a metasurface device with polarization switchable image recognition function based on diffraction neural network and dynamic phase combined with geometric phase. Background Art

[0002] In the 17th century, Dutch physicist Christian Huygens first proposed the Huygens principle. Huygens believed that each point on the wavefront can be regarded as a source of secondary spherical sub-waves, and these secondary spherical waves will superimpose on each other to produce new wavefronts. In the 18th century, Augustin Fresnel made further improvements on the theory proposed by Huygens, that is, when the secondary spherical waves emitted by each point on the wavefront are coherent, these coherent lights can superimpose new light fields at new positions, producing diffraction, interference and other phenomena. Optical neural networks are expanded on this basis.

[0003] All-optical deep learning is an emerging research field that shows great potential to change the limitations of traditional image recognition systems. Traditional image recognition systems rely on electronic computers to process large amounts of data and execute complex algorithms, which have limitations in resource consumption and speed. In 2018, Ozcan's research group first disclosed an artificial neural network technology based on optical devices, namely the diffractive neural network (DNN). 2 NN), which uses the principle of optical diffraction for information processing. 2 In NN, the units on each layer of the metasurface are similar to artificial neurons, and the connections between them are realized through the Huygens-Fresnel diffraction law. When light passes through these metasurfaces, it is modulated by amplitude and phase. Through computational optimization, the response of each metasurface unit can be adjusted to make D 2 NN can perform functions such as image recognition, classification, and eigenvalue problem solving. The technology operates at the speed of light, has low energy consumption and high parallel processing capabilities, and therefore has obvious advantages in data processing speed and energy efficiency. Compared with traditional electronic computers, D 2 NN does not require complex circuits and a lot of electricity, and can process large amounts of data in a shorter time. Therefore, all-optical deep learning technology shows broad prospects in image recognition and other applications that require efficient data processing.

[0004] Metasurface technology is a structure composed of a subwavelength-scale array of nanounits that can precisely manipulate light. These nanounits can locally control the phase, amplitude, and polarization state of the incident light. The dynamic phase, also known as the propagation phase, is a concept that describes the phase change of electromagnetic waves as they propagate in space. The transmission of light in the material produces phase accumulation, and the phase of each unit result is controlled by changing the size of the structure. Assuming that the incident wavelength is λ, the refractive index of the material is n, and the distance the light travels in the material is d, then the phase can be written as: Φ = nk0d. Where k0 = 2π / λ is the free space wave vector. In order to meet the functional requirements, we need to select a suitable phase distribution unit structure. Metasurface devices with dynamic phase regulation usually require a large number of structural parameters as a database. In order to design a unit structure with a more accurate phase distribution, the superstructure is parameter scanned. For example, rectangular nanorods of different sizes can be scanned using FDTD electromagnetic simulation software to obtain the corresponding dynamic phase and transmittance. Based on the database of the relationship between the dynamic phase and nanorods of different sizes, the appropriate size structure can be finally selected according to the design target phase. This design method is relatively simple, but for the design of complex functions, it may be necessary to scan more sized structures to match the phase distribution. In general, this method provides an effective way to design metasurfaces, and is often combined with geometric phase to achieve more novel functions. Geometric phase, also known as Pancharatnam-Berry (PB) phase, was first proposed by British professor Berry, using the different rotation angles of the antenna at each position to control the phase. Simply put, from the initial state of the Poincaré sphere to the final state, and then back to the initial state, there will be a cross phase change, and the value of this change is twice the antenna rotation angle. By designing and selecting the appropriate size and shape of the unit, arranging them according to the phase distribution calculated in the diffraction neural network, and introducing the special effect of geometric phase, the light obtains a phase delay related to its rotation angle when passing through these structures, which can directly realize information extraction and processing in the optical domain, and provide the possibility for the development of compact and high-performance image recognition equipment.

[0005] Traditional metasurfaces can usually only achieve asymmetric phase control, which limits their potential in some applications, especially the limitations of traditional metasurfaces in polarization control, which seriously weakens their application potential in directional polarization control. This means that in scenarios where light waves with a specific polarization state are required for highly directional transmission, it is difficult for traditional metasurfaces to achieve optimal performance, cannot adapt to different application requirements and environmental changes, and cannot simultaneously achieve the confidentiality of unidirectional transmission and the sharing of bidirectional transmission, but only has a single function. Summary of the invention

[0006] In view of the problem of single function of current metasurface devices, in order to enhance the ability of flexible and switchable image recognition, a metasurface device with polarization switchable image recognition function and its implementation method are proposed, which can identify and classify according to different incident linear polarized light on both sides of the metasurface device, achieving a good balance between the confidentiality of unidirectional transmission and the sharing of bidirectional transmission, which not only broadens the application scope of metasurface devices, but also brings new opportunities to the field of information processing. A flexible security framework is constructed by the confidentiality capability achieved by unidirectional transmission and the information sharing capability provided by bidirectional transmission. Users can exchange information securely with other users without leaking sensitive information. In a multi-user or multi-device environment, the bidirectional transmission capability enables different users to share and exchange information with each other without changing the device settings. This feature greatly facilitates effective communication in collaborative work and shared applications.

[0007] The diffraction neural network of the present invention is used for target recognition and is composed of a multi-layer structure. The phase distribution of each layer is precisely designed to modulate the incident light wave. The parameters of the diffraction layer are optimized by using a gradient descent algorithm and a back-propagation algorithm to achieve the image recognition function. The dynamic phase is combined with the geometric phase principle to generate a metasurface structure array according to the optimized diffraction layer phase parameters to achieve the corresponding function. The metasurface device can realize polarization switching of the image recognition function, greatly improving its functionality. It has the advantages of small size, high integration, low computing resource consumption, and short training time. It is suitable for wide applications in the fields of intelligent perception and artificial intelligence.

[0008] The technical solution of the present invention is:

[0009] A metasurface device with polarization-switchable image recognition function, comprising a metal digital imaging plate, a metasurface layer and a detection plane, the three parts being arranged in sequence to form a complete image recognition system; the metal digital imaging plate comprises a transparent silicon dioxide substrate, a gold film placed on the silicon dioxide substrate, and the shape is a pattern to be recognized, which plays a role in optical amplitude encoding; the metasurface layer has a periodic structure, comprising a silicon material dielectric substrate, a silicon column structure array placed on the silicon material dielectric substrate and a metal grating structure on the back of the substrate; these silicon column structures can diffract incident light according to a specific phase distribution and rotate the incident light at a 45-degree angle based on a geometric phase rotation, and realize a polarization-switchable image recognition function based on a neural network through the filtering effect of the metal grating; the phase distribution is obtained through the training result of the diffraction neural network model, so as to accurately control the light;

[0010] The structure of the diffraction neural network includes an input layer, a diffraction layer, and an output layer: in the input layer, the pattern to be recognized will be encoded into the amplitude of the input layer of the diffraction neural network, so that the image information is converted into a light signal, which is then transmitted to the diffraction layer for processing; the diffraction layer is constructed by a finely adjusted phase distribution, similar to the weight matrix in a deep learning neural network, which modulates the propagation of the light wave by precisely controlling the phase of the incident light wave; the function of the output layer is to detect the light wave output by the diffraction layer and complete the image recognition task; this output layer is a pre-set detection plane, and the diffraction network can be trained to map the input target to the corresponding detection area; a number of detection areas will be divided on this detection plane, and the number of sub-detection areas matches the number of categories of the object to be recognized;

[0011] The metasurface device realizes diffraction of incident light of a specific frequency after passing through the metal digital imaging plate, and finally forms a diffraction spot at the preset corresponding position of the detection plane, realizing the image recognition function based on the diffraction neural network polarization switchability; extracting the optimized diffraction layer phase distribution text file from the diffraction neural network structure, applying the dynamic phase combined with the geometric phase principle to construct a metasurface silicon column array, and selecting eight silicon column structures with different responses to linear polarized light by scanning the silicon column parameters to ensure that each of them has a quarter-wave plate function at the set frequency and achieves phase coverage from 0 to 2π. After realizing the phase control function, it also has the function of rotating all silicon columns 45 degrees to realize polarization conversion, and cooperates with the filtering function of the metal grating structure on the back of the substrate, similar to the role of a polarizer, allowing y-polarized electromagnetic waves to pass through, while blocking x-polarized terahertz waves from passing through; combining these two layers, when x-polarized light is incident from the front, it first passes through the metasurface layer, forming a left-handed circularly polarized diffraction spot, and then passing through the grating layer, because the grating layer only allows y-polarized light to pass through, a y-polarized diffraction spot is formed on the detection plane; on the contrary, when the x-polarized light is incident from the back, it first passes through the grating layer, and the grating layer reflects all the x-polarized light, so no diffraction spot can be formed on the detection plane, so the image cannot be recognized under the x-polarized light in the reverse direction; similarly, when the y-polarized light is incident from the front, it first passes through the metasurface layer to form a right-handed circularly polarized diffraction spot, and then passes through the grating layer, because the grating layer only allows y-polarized light to pass through, a y-polarized diffraction spot is formed on the detection plane; when the y-polarized light is incident from the back, it first passes through the metal grating, the y-polarized light can pass through the grating, and then passes through the metasurface layer to form a left-handed circularly polarized diffraction spot on the detection plane, so bidirectional image recognition can be realized in the y-polarization state, thereby realizing the polarization-switchable image recognition function;

[0012] First, the incident light is input and passes through a metal digital imaging plate placed on a silicon dioxide substrate. The metal part of the imaging plate is opaque, while the pattern hollow part is transmissive. The transmitted light is then irradiated onto the metasurface layer. During the forward irradiation process, the transmitted light is first diffracted and polarized by the silicon pillar structure array placed on the dielectric substrate. After being filtered by the metal grating, a diffraction spot is finally formed on a detection plane at a certain distance from the metasurface. During the forward transmission process, the x-polarized light is converted into left-handed circularly polarized light by the silicon pillar array structure. After passing through the grating, a diffraction spot of the y-polarization state is formed on the detection plane, corresponding to the classification of the input image. The y-polarized light is converted into right-handed circularly polarized light by the silicon pillar array structure. After passing through the grating, a diffraction spot of the y-polarization state is formed on the detection plane, corresponding to the classification of the input image.

[0013] During the reverse irradiation process, the transmitted light is first filtered by the metal grating, and then diffracted and polarized by the silicon pillar structure array placed on the dielectric substrate, and finally forms a diffraction spot on the detection plane at a certain distance from the metasurface; during the reverse transmission process, the x-polarized light is first affected by the metal grating. Since the grating along the y direction will reflect the x-polarized light, almost no light can reach the silicon pillar array structure, and therefore no diffraction spot can be formed; while the y-polarized light can pass through the grating and then through the silicon pillar array structure to be converted into left-handed circularly polarized light, forming a diffraction spot on the detection plane. The position of the diffraction spot corresponds to the input image category.

[0014] Furthermore, the silicon pillar structure is designed according to the transmittance requirements and can operate at different light wavelengths.

[0015] The preferred metal digital imaging plate is a layer of gold film coated on a silicon dioxide substrate by vacuum coating technology, the thickness of the silicon dioxide substrate is 500 microns, and the thickness of the gold film is 5 microns.

[0016] Preferably, the dielectric substrate is made of silicon material with a thickness of 500 microns and a structural length and width of 13,000 microns.

[0017] Furthermore, the incident light uses two types of linear polarized light in the x direction and the y direction.

[0018] A method for implementing polarization-switchable image recognition based on a diffractive neural network is implemented based on a metasurface device with polarization-switchable image recognition function, comprising the following steps:

[0019] The diffraction neural network consists of three parts: input layer, diffraction layer, and output layer. The gradient descent algorithm and back propagation algorithm are used to train and optimize the diffraction layer parameters. The recognition task data set is used to train the diffraction neural network diffraction layer parameters and determine the diffraction layer phase distribution. The optimized diffraction layer phase distribution text file is extracted, and the metasurface rectangular silicon column array is constructed using the dynamic phase combined with the geometric phase principle. The designed silicon column realizes the function of a quarter-wave plate at a specific frequency point, that is, the phase difference of the light wave in the fast axis and the slow axis direction is 90 degrees. This phase difference will change the polarization state of the light. Specifically, when the linearly polarized light is at an angle of 45 degrees When incident on the optical axis of the quarter-wave plate, the light wave is decomposed into components along the fast axis and the slow axis. These two components will have a phase difference of 90 degrees after passing through the wave plate, thereby combining into circularly polarized light; the polarization conversion function is realized by rotating each silicon column in the silicon column array by 45 degrees; thereby generating a metasurface structure file; at the same time, the pattern to be recognized is processed, the pattern is hollowed out using a gold film, a silicon dioxide substrate is added, and the distance from the metasurface layer is set, and the substrate is placed parallel to the metasurface layer. The irradiation light source passes through the metal imaging plate and the metasurface layer in turn, realizing the design of a metasurface device with polarization-switchable image recognition function;

[0020] The input layer of the diffraction neural network encodes the image information to be recognized and converts it into optical signals; these optical signals carry the amplitude distribution information of the image shape and are input to the next diffraction layer for processing; the optical signal is transmitted from the input layer to each diffraction layer, and the layers are interconnected through neurons to simulate the amplitude of the modulated light beam; in the diffraction layer, the optical signal interacts with these layers layer by layer and is processed through a multi-layer diffraction structure; each diffraction layer is modulated according to the optical signal transmitted by the previous layer to extract and enhance the characteristic information of the image; the output layer is responsible for detecting the light waves processed by all diffraction layers and performing image recognition tasks; in order to achieve the best target recognition effect, the diffraction layer obtains the optimal diffraction layer phase parameters through training; these parameters are applied to the distribution of the silicon pillar structure array to form a metasurface layer with polarization switchable function; the metasurface device can ultimately achieve efficient image recognition function under different polarization states, thereby improving the accuracy and efficiency of image recognition;

[0021] The specific training process of the diffractive neural network model is as follows:

[0022] Step 1) In the diffraction neural network model, the phase and amplitude changes of the light wave are calculated according to the Rayleigh-Sommerfeld diffraction integral formula, and the network model determines the bias term through the transmission coefficient of each layer of the metasurface; the model uses the forward propagation process to guide the incident light wave carrying the target information through a multi-layer metasurface structure with a specific configuration to form a diffraction pattern reflecting the characteristics of the target object on the detection plane; the Rayleigh-Sommerfeld diffraction theory describes the propagation mode of the light field from the nth layer to the n+1th layer, where represents the transmission coefficient of the nth layer, where A is the amplitude, is the phase, i is the imaginary unit, and a=1 is set when only the pure phase type diffraction neural network is trained; the light field distribution of the propagation form of Rayleigh-Sommerfeld diffraction theory is expressed as:

[0023]

[0024] Among them, U(r n ) is the transmitted electric field distribution of the previous layer, U(r n+1 ) is the transmission electric field distribution on the n+1th layer, n is the number of diffraction layers, and in the formula h(r n+1 -r n ) represents the impulse response during the transmission process, and the expression is:

[0025]

[0026] in, is the point r on the nth layer. n To point r on the n+1th layer n+1 The Euclidean distance between them, λ is the incident wavelength, is the imaginary unit, x n+1 ,y n +1 、z n+1 are the spatial coordinate components of a point at the n+1th layer, are the spatial coordinate components of the ath neuron in the nth layer respectively; S represents the integral area, which is all the points in the entire wavefront area; the integral represents the weighted and superimposed light field distribution of the entire nth layer to calculate the light field distribution of a point on the n+1th layer; this method can be used to predict and calculate the propagation of light waves in media with different layer structures; the number of diffraction layers is represented by n, where a single diffraction layer n=1 is used, and the n+1th layer is used to represent the detection plane, and according to the energy distribution S of the sub-detection area of ​​the detection plane j =|U j | 2 , j is the sub-detector region number, and the recognition result is determined by the category corresponding to the region with the highest energy;

[0027] Step 2) A commonly used image classification dataset is selected as a training dataset, and the images to be identified and classified are selected as the training set. The purpose is to train the diffraction layer phase parameters of the diffraction neural network to maximize the energy distribution of the corresponding category sub-detection area and minimize the energy distribution of other sub-areas; in the optimization process, the mean square error loss function MSE is used to evaluate the difference between the energy distribution of different sub-detector areas and the target energy distribution, and the number of identified targets is represented by C:

[0028]

[0029] where s c represents the energy distribution received in the final detection area, g c represents the target energy distribution, represents the error between the target area and the detection area; based on the obtained diffraction pattern and its deviation from the expected result, a loss function is defined to evaluate the recognition accuracy of the diffraction neural network under the current parameters;

[0030] Step 3) Apply the back propagation module and combine it with the gradient descent algorithm to calculate the gradient of the loss function with respect to the phase parameter of the diffraction layer to guide the optimization process of the phase distribution; after multiple iterations until Stop training when convergence occurs and the accuracy cannot be improved; repeat the above steps to adjust the number of diffraction layers to find the number of diffraction layers that can achieve the object recognition accuracy required by the designer;

[0031] Step 4) iteratively executes step 1) to step 3), continuously updating the diffraction layer phase parameter until a predetermined recognition accuracy or other performance index is met, and the training and optimization of the diffraction neural network is completed;

[0032] According to the phase distribution of the diffraction layer determined in step 2), eight rectangular silicon pillar structures of different sizes are selected, which have a quarter-wave plate function at a set frequency point. Rotating them 45 degrees can realize the function of polarization conversion, and they are arranged in an array according to the calculated phase distribution to form a multi-layer metasurface; the unit structure parameters of these metasurfaces include the length L, width W, height H and period length P of the silicon pillar; the performance of silicon pillar structures of different sizes under linear polarized light irradiation is calculated to select the optimal parameter combination required at a specific wavelength and determine the size of the selected silicon pillar; according to the geometric phase principle, the expected phase distribution is achieved by rotating the silicon pillar structure; finally, the silicon pillar structure is arranged according to the phase distribution obtained by training the diffraction layer of the diffraction neural network, and a metasurface device with polarization-switchable image recognition function is constructed based on this;

[0033] The final phase distribution is in is the phase obtained by training the all-optical diffraction neural network;

[0034] This method can be applied to the recognition tasks of various data sets. By calculating the phase distribution of the diffraction layer and applying it to the metasurface layer, the metasurface layer consists of a double-layer structure to realize the image classification function of polarization switching. The silicon pillar array combines the geometric phase with the propagation phase, and has the ability of phase regulation and polarization conversion. It is composed of silicon pillars rotated 45° and of different sizes and arranged periodically to meet the phase and amplitude conditions of the quarter-wave plate. The second layer is a metal grating with its long axis along the x-direction, similar to a polarizer, which allows y-polarized terahertz waves to pass through and blocks x-polarized waves. Combining these two layers, when x-polarized light is incident on the front, It first passes through the metasurface layer to form left-handed circularly polarized light, and then passes through the grating layer to allow only y-polarized light to pass through, forming a y-polarized diffraction spot on the detection plane; conversely, when x-polarized light is incident from the back, the grating layer reflects the x-polarized light, and no spot is formed on the detection plane, thereby realizing the directional image recognition function; similarly, when y-polarized light is incident from the front, the metasurface layer forms right-handed circularly polarized light, and after passing through the grating layer, a y-polarized diffraction spot is also formed on the detection plane; and when y-polarized light is incident from the back, it first passes through the grating and then passes through the metasurface layer, generating a left-handed circularly polarized diffraction spot on the detection plane, thereby realizing the polarization conversion function.

[0035] Furthermore, the diffraction neural network training process uses a single diffraction layer, sets the distance from the input layer to the diffraction layer to 10,000 microns, the distance from the diffraction layer to the output layer to 10,000 microns, the learning rate to 0.001, the size of the neuron to 130 microns, and in order to save training time, eight samples are taken as a batch and input into the neural network.

[0036] Furthermore, the training model software uses the TensorFlow framework, which is based on the Python programming language.

[0037] Furthermore, electromagnetic full-wave simulation software can be used to calculate the performance of silicon pillar structures of different sizes under linear polarized light illumination.

[0038] The beneficial effects of the present invention are:

[0039] 1. The present invention discloses a metasurface device with polarization-switchable image recognition function based on a diffraction neural network and a combination of a dynamic phase and a geometric phase. It is the first metasurface device with asymmetric transmission and polarization-switchable image recognition function realized in the terahertz band. This method can not only utilize a diffraction neural network for image recognition, but also achieve different recognition effects on both sides of the metasurface according to different linear polarized light, thereby improving the efficiency and accuracy of the recognition task.

[0040] 2. The metasurface device disclosed in the present invention realizes polarization-switchable image recognition function based on diffraction neural network and dynamic phase combined with geometric phase. The ability to operate in different bands such as terahertz, near-infrared and microwave is taken into consideration during design, and its scalability and flexibility in multi-band operation.

[0041] 3. Compared with traditional image recognition devices, the proposed metasurface device with polarization-switchable image recognition function based on diffraction neural network and dynamic phase combined with geometric phase has the advantages of low energy consumption, fast response speed, small size and high integration; the system does not involve mechanical components and can operate at the speed of light, which greatly improves the computing efficiency and response speed; the metasurface silicon pillar has a simple structure, a wide range of materials, is easy to process and has a wide range of applications; the proposed metasurface device with polarization-switchable image recognition function based on diffraction neural network and dynamic phase combined with geometric phase can achieve frequency band conversion, multi-frequency point conversion and multi-target recognition by adjusting structural size parameters, setting response frequency bands and other methods according to actual application scenarios and requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flow chart of the calculation method for designing the metasurface device of the present invention;

[0043] Figure 2 Schematic diagram of the positions of diffraction spots of different inputs and corresponding output detection layers of the metasurface device of the present invention;

[0044] Figure 3 Parameter scanning result diagram of the dynamic phase design of the metasurface device of the present invention; (a) schematic diagram of the unit structure of the transmission phase; (b) transmission phase distribution when x-polarized light is incident; (c) transmission phase distribution when y-polarized light is incident; (d) transmittance distribution when x-polarized light is incident; (e) transmittance distribution when y-polarized light is incident;

[0045] Figure 4 Transmittance and phase distribution data diagram of 8 selected rectangular silicon pillars for x-polarized incident and y-polarized incident for the dynamic phase design of the metasurface device of the present invention, and the selected 8 quarter-wave plates simultaneously cover 0 to 2π;

[0046] Figure 5 is a top view of a single periodic structure of the metasurface device of the present invention, wherein: Figure 5 (a) and (b) are schematic diagrams of the unit periodic structure without rotation and the unit periodic structure with rotation, respectively. Angle diagram, Figure 5 (c) is the grating structure on the back side of a single periodic structure of the metasurface device, where the gratings are arranged along the y polarization direction; Figure 5(d) is the main view of a single periodic structure of the metasurface device. The unit structure is divided into three layers, consisting of a silicon pillar on the upper layer, a silicon substrate in the middle, and a metal grating on the bottom layer;

[0047] Figure 6 A top view of sixty-four periodic structures of the metasurface device of the present invention;

[0048] Figure 7 Schematic diagram of the input metal digital imaging plate of the metasurface device of the present invention;

[0049] Figure 8 A schematic diagram of the polarization switchable recognition function of the metasurface device of the present invention;

[0050] Fig. 9 The image and simulation effect of the switchable polarization handwritten digital image recognition in the embodiment of the present invention.

[0051] 1. Silicon pillar structure array; 2. Dielectric substrate; 3. Metal grating; 4. Metal imaging plate; 5. Silicon dioxide substrate. DETAILED DESCRIPTION

[0052] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0053] The present invention proposes a new structure of a double-layer metasurface combined with a grating, which can realize asymmetric polarization control of electromagnetic waves in different propagation directions. The advantages of the dynamic phase and geometric phase of this metasurface are combined with the diffraction neural network, and the micro-nano unit structure of the metasurface is used to simulate neurons in deep learning, thereby obtaining a polarization-switchable bidirectional image recognition functional device. This optical neural network significantly improves the data processing capability and brings breakthrough progress to all-optical computing and machine vision. Since photons are extremely fast during propagation, this network can process large amounts of data in a very short time, greatly improving computing efficiency. In addition, this integrated design maintains the miniaturization and integration characteristics of the system, making it more convenient to apply to practical scenarios. This lays a solid foundation for processing more complex optical signals and realizing intelligent sensing applications.

[0054] Diffractive neural networks are computational models that emulate the principle of diffraction in optics. In this model, light waves act as carriers of information, transmitted between layers of neurons in a medium. The calculations of each layer are supported by the Huygens-Fresnel principle, that is, the propagation of light waves can be understood as the vibration of the wavefront at each point. When light waves flow from one level to another, secondary waves are generated, thereby achieving top-down signal transmission. These different diffraction layers can be transmitted or reflected, and each point is like a neuron, participating in information processing. The diffraction patterns generated when light waves propagate between different media are used to process information. The training of the network relies on deep learning on a computer, including multiple rounds of training and parameter adjustment to optimize the network structure and improve performance.

[0055] The present invention discloses a metasurface device based on a diffraction neural network and a combination of dynamic phase and geometric phase, which can realize polarization switchable image recognition function. The phase distribution of the metasurface device is calculated by a diffraction neural network, which consists of three parts: an input layer, a diffraction layer, and an output layer. The diffraction layer is precisely calculated and designed to simulate the weight matrix function in a deep learning neural network and is optimized and adjusted according to specific tasks. The diffraction layer parameters are trained using gradient descent and back propagation algorithms, and the optimal diffraction layer phase distribution is determined in combination with a specific data set to realize the polarization switching function of image recognition.

[0056] The metasurface of the present invention has a single periodic structure, including a dielectric substrate and a silicon column array structure for phase control. The parameters are scanned and the appropriate geometric dimensions of the metasurface silicon column structure are selected therefrom. The selection conditions meet the conditions that the transmittance is greater than 0.75, the phase difference under x and y polarized light irradiation is half π, and the phase covers 0-2π. The direction of the metasurface silicon column is rotated 45 degrees using the geometric phase principle. According to the optimized diffraction layer phase distribution, the structure file of the metasurface is generated to ensure that the diffraction effect best matches the image recognition requirements and realizes strict modulation of the incident light beam phase. A metal digital imaging plate is illuminated with plane light to obtain a light beam carrying the target image information. When the light beam is projected onto the metasurface, it will follow a predefined path to generate a specific diffraction pattern and form a diffraction spot on the output plane. The metasurface device developed by the present invention has the function of switchable recognition under different polarized light incidence, which improves the efficiency and accuracy of the target recognition task.

[0057] The metasurface silicon pillar is made of silicon material, with a unit structure period of 130 microns and a thickness of 500 microns. It functions as a quarter wave plate at a set working frequency of 0.6 THz, and the number of arrays is 100*100. The silicon pillar can also be made of metal material.

[0058] The dielectric substrate is made of silicon material with a thickness of 500 microns and a structure length and width of 13,000 microns.

[0059] The metal digital imaging plate is made of gold material with a thickness of 5 microns, and its substrate is silicon dioxide with a thickness of 500 microns.

[0060] A design of a metasurface device with polarization-switchable image recognition function based on diffraction neural network and dynamic phase combined with geometric phase control. The device simulates functions similar to traditional deep learning networks by designing a specific metasurface structure, but processes in the optical domain, thereby achieving fast and efficient target detection and classification, belonging to the fields of optical neural networks, micro-nano optics, optical information processing and machine vision application technology.

[0061] The diffraction neural network architecture of the metasurface device that realizes polarization-switchable image recognition function based on diffraction neural network and dynamic phase combined with geometric phase control consists of three parts: input layer, diffraction layer, and output layer. The input layer uses a digital coding imaging plate to encode the image information to be identified, and applies the image information encoding to the beam amplitude channel as information input to the diffraction layer; the diffraction layer is composed of phase distribution, simulates the weight matrix in the deep learning neural network, and is physically realized by the metasurface micro-nano structure; the geometric phase control mechanism is applied on the diffraction layer, and is composed of a computationally designed silicon column structure for precise modulation of the beam phase distribution; an output layer is used to detect the light wave output by the diffraction layer and perform image recognition and spatial positioning. The output layer is a pre-prepared detection plane. The recognition object used in this embodiment is a handwritten digital pattern. There are four handwritten numbers to achieve four classifications, and four equal-sized sub-detection areas are delineated on the detection plane.

[0062] Among them, the diffraction neural network calculates the entire device in an optical computing way, and its output layer is an optical signal.

[0063] like Figure 1 As shown, the design process method specifically implemented in this embodiment is as follows:

[0064] Handwritten digits are used as training sets and test sets to train the all-optical diffraction neural network. The gradient descent algorithm and the back propagation algorithm are used in combination with a specific recognition task data set to train the diffraction layer parameters in the diffraction neural network and obtain the diffraction layer phase distribution of the diffraction neural network. Based on the determined diffraction layer phase distribution, the silicon pillar array on the metasurface is designed through the concept of dynamic phase combined with geometric phase, and the corresponding metasurface structure engineering design file is created to establish a simulation model. According to the principle of combining dynamic phase and geometric phase with the diffraction neural network based on the metasurface, a metasurface device that can realize image recognition under different polarization conditions is obtained, which promotes the miniaturization and functional integration of image recognition and classification equipment.

[0065] The Diffractive Deep Neural Network is an all-optical deep learning framework, whose core is composed of a series of multi-layered diffractive surfaces working together. These surfaces physically form a neural network that can perform statistical learning of complex functions optically. The calculation and prediction mechanism of the physical network is all-optical, and the learning part of the design is done by computer. This example verifies the D 2 The effectiveness and reasoning ability of NN (Diffractive Deep Neural Network) are used to realize a metasurface device that can switch image recognition under different polarization conditions. The simulation software is based on the finite difference time domain method.

[0066] The examples used are x-polarized light and y-polarized light at 0.6 THz;

[0067] like Figure 2 As shown, the input layer uses a metal imaging plate to realize the information encoding of the object image to be recognized, and encodes the image information into the beam amplitude channel. The detection plane of the handwritten digits is divided into 4 discrete areas, representing the handwritten digits "2", "3", "6", and "7" respectively.

[0068] like Figure 3 As shown, the metasurface device is composed of multiple periodic structure arrays based on dynamic phase. In order to enable the unit structure to achieve phase coverage from 0° to 180°, the rectangular column made of high-resistance silicon is subjected to parameter scanning to obtain the corresponding transmission phase and transmittance of the rectangular structure when the length and width are 30-120μm respectively. Finally, the size of the rectangular silicon column is selected according to the target phase distribution.

[0069] like Figure 3 , 4 , 5 show the transmission ratio and phase diagram of the y-component and x-component of the eight selected structures; the metasurface device is composed of multiple periods and is designed as a 100*100 period array structure. The FDTD simulation software is used to perform parameter scanning at 0.6THz to find the optimal structural parameters, and eight rectangular silicon pillars with different lengths and widths are found. The size data are given in Table 1. The thickness of the rectangular silicon pillar is h1=500μm, the substrate thickness is h2=500μm, and the total thickness is h 1+h2, the unit period length p in the x and y directions is 130μm (the length and width of the substrate are both p, and the substrate of each periodic structure is a square), and the incident wavelength is 500nm. At 0.6THz, the same polarization transmittance of the x-polarized light and the same polarization transmittance of the 8 selected silicon pillars are the same, and the phase difference between the transmitted x and y polarizations is approximately 90°. The selected structure is close to a quarter wave plate. In addition, these 8 silicon pillars meet the phase coverage of 0 to 2π. In order to realize the polarization conversion function, the rotation angle θ is set to 45 degrees, which can convert the incident linearly polarized light into circularly polarized light. A gold material grating is used on the back of the dielectric substrate. The grating is placed along the y-axis with a thickness of h au =10μm, interval d g =12.5μm, grating width w g =20μm. There are four grating structures in each unit cell period.

[0070] Table 1: Length and width data of 8 selected rectangular silicon pillars for the dynamic phase design of the metasurface device of the present invention

[0071]

[0072] like Figure 5 As shown, (a) top view of the unit structure; (b) silicon pillar rotates around the center of the unit structure at an angle of θ; (c) schematic diagram of the back grating of the unit structure; (d) front view of the unit structure, the overall unit structure is divided into three layers; the long axis of the silicon pillar rotates around the center of the unit periodic structure, and the angle between it and the x-axis of the substrate is θ (formed by the rotation angle of the silicon pillar structure), which is 45 degrees here. When the incident light is linearly polarized light, the transmitted light is circularly polarized light.

[0073] like Figure 6 As shown, a total of sixty-four periodic structure diagrams of the partial area are displayed (the overall structure is 100*100).

[0074] like Figure 7 As shown, one of the metal imaging plates is selected for display. The side length is 100 times that of the unit periodic structure, that is, 13,000 microns, to achieve the regulation of the amplitude of the incident linear polarization light and apply the image information encoding to the beam amplitude channel. The optical digital imaging plate of the metasurface device uses a patterned gold film with a thickness of h3 = 5 microns, placed on a silicon dioxide substrate with a thickness of h4 = 500 microns, and the number of silicon pillar structure arrays used is 100*100. Then the length and width of the metal digital imaging plate are both 100p, and the overall position is located 10,000 microns from the upper surface of the metasurface layer.

[0075] like Figure 8As shown in the figure, the working principle of the designed polarization-switchable metasurface device. When x-polarized light is incident from the forward direction of the metasurface, it first acts on the silicon pillar structure after passing through the metal imaging plate. The silicon pillar has the function of a quarter-wave plate and rotates 45° to achieve the function of polarization conversion. Then it passes through the metal grating layer with the long axis along the x direction, which acts like a polarizer. The terahertz wave with y polarization is allowed to pass through, while the terahertz wave with x polarization is blocked. Therefore, a y-polarized diffraction spot is formed on the detection plane; on the contrary, when the x-polarized light is irradiated in the reverse direction, it first interacts with the grating, causing the light to be reflected, so the input image cannot be recognized. Similarly, when the y-polarized light is incident from the front, it first passes through the metasurface layer to become right-handed circularly polarized light, and then passes through the grating layer. Since the grating layer only allows y-polarized light to pass through, a diffraction spot with y polarization state is formed on the detection plane. When y-polarized light is incident from the back, it first passes through the metal grating and then undergoes the image recognition and polarization conversion action of the metasurface layer, forming a diffraction spot in a left-handed circular polarization state on the detection plane, thereby realizing the image recognition function with polarization switchable function.

[0076] like Fig. 9 As shown in the figure, in the classification simulation, the handwritten digit classification component has undergone 20 generations of iterative training, and the simulation test accuracy has reached more than 92%. A total of 24,272 handwritten digit images are used as training data sets, and 4,028 handwritten digit images are selected as test sets. The simulation results are good, indicating that the design theory is effective and the device can correctly recognize handwritten digits. Four digital combination patterns to be recognized are selected for demonstration. When the input number is 2, there is a bright diffraction spot in the upper left corner of the four preset boxes in the detection area behind the metasurface; when the input number changes to 3, there is a bright diffraction spot in the upper right corner of the four preset boxes; when the input number changes to 6, there is a bright diffraction spot in the lower left corner; and when the input number changes to 7, there is a bright diffraction spot in the lower right corner of the four preset boxes in the detection area behind the metasurface. When x-polarized light is input in the forward direction, the input image can be recognized and classified, and the reverse input cannot recognize the image result; in the case of y-polarized light input, both forward and reverse inputs can recognize and classify the input image.

[0077] In this embodiment, all-silicon material is selected as the dielectric substrate and the super-surface silicon column structure, and the silicon column structure is obtained by using photolithography and magnetron sputtering coating technology. The present invention is based on the diffraction neural network and dynamic phase combined with geometric phase to realize polarization switchable image recognition of the super-surface device. At a frequency of 0.6THz, the x and y linear polarized light are respectively irradiated from the forward and reverse directions to the super-surface layer after passing through the digital imaging plate, and the image recognition function of polarization switching of the input digital can be realized, and the recognition result is presented on a plane 10,000 microns away from the other side of the super-surface layer.

[0078] The present invention aims to achieve the design and optimization of a metasurface device that constructs a diffraction neural network and a dynamic phase combined with a geometric phase to achieve polarization-switchable image recognition through a diffraction neural network. As a cutting-edge means of applying deep learning algorithms to optical wavefront design, the diffraction neural network fully considers the physical constraints in the optical system, while using the powerful fitting ability of the neural network to explore the optimal silicon column arrangement. The phase distribution of the diffraction layer is obtained according to the metasurface-based diffraction neural network optimization method, and the silicon column structure corresponding to the phase modulation is selected according to the phase distribution. Then, a single-layer metasurface is prepared, and finally the diffraction neural network and the dynamic phase combined with the geometric phase are successfully obtained to realize a metasurface device with polarization-switching image recognition function.

[0079] The above-mentioned embodiment only expresses one implementation mode of the present invention, and its description is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A polarization-switchable metasurface device with image recognition function, characterized in that: The metal digital imaging plate includes a metal digital imaging plate, a metasurface layer and a detection plane, which are arranged in sequence to form a complete image recognition system; the metal digital imaging plate includes a transparent silicon dioxide substrate and a gold film placed on the silicon dioxide substrate, and the shape is the pattern to be recognized, which plays the role of optical amplitude coding; The super surface layer has a periodic structure, including a silicon material dielectric substrate, a rectangular silicon pillar structure array placed on the silicon material dielectric substrate, and a metal grating structure on the back of the substrate; These rectangular silicon pillar structure arrays can diffract incident light according to a specific phase distribution and rotate the polarization of the incident light based on a 45-degree geometric phase rotation. The combination of the filtering effect of the metal grating structure realizes the polarization switchable image recognition function based on the diffraction neural network; the phase distribution is obtained through the training results of the diffraction neural network model to precisely control the light; The structure of the diffraction neural network includes an input layer, a diffraction layer, and an output layer: in the input layer, the pattern to be identified is encoded into the amplitude of the diffraction neural network input layer, thus converting the image information into a light signal, which is then transmitted to the diffraction layer for processing; the diffraction layer is constructed by a finely tuned phase distribution, similar to the weight matrix in the deep learning diffraction neural network, which modulates the propagation of the light wave by precisely controlling the phase of the incident light wave; the output layer is used to detect the light wave output by the diffraction layer and complete the image recognition task; this output layer is a pre-set detection plane, and the diffraction neural network can be trained to map the input target to the corresponding detection area; several detection areas will be divided on this detection plane, and the number of sub-detection areas matches the number of categories of the object to be identified; The metasurface device realizes diffraction of incident light of a specific frequency after passing through the metal digital imaging plate, and finally forms a diffraction spot at the preset corresponding position of the detection plane, realizing the polarization-switchable image recognition function based on the diffraction neural network; extracts the optimized diffraction layer phase distribution text file from the diffraction neural network structure, and constructs a rectangular silicon pillar structure array by applying the dynamic phase combined with the geometric phase principle. Eight rectangular silicon pillar structures with different responses to linear polarized light are selected by scanning the parameters of the rectangular silicon pillar structures in the rectangular silicon pillar structure array to ensure that each of them has a quarter-wave plate function at the set frequency and achieves phase coverage from 0 to 2π. After realizing the phase control function, all the rectangular silicon pillar structures are rotated 45 degrees to realize the polarization conversion function, and cooperate with the filtering function of the metal grating structure on the back of the substrate, similar to the role of a polarizer, allowing y Polarized electromagnetic waves pass through, but x Polarized terahertz waves pass through; the rectangular silicon pillar structure array and the metal grating structure are combined. x When polarized light is incident from the front, it first passes through the rectangular silicon pillar structure array to form a left-handed circularly polarized diffraction spot, and then passes through the metal grating structure. y Polarized light is transmitted, thus forming a y Polarized diffraction spot; on the contrary, when x When polarized light is incident from the back, it first passes through the metal grating structure, which reflects all the x-polarized light, so no diffraction spot can be formed on the detection plane. x The image cannot be recognized when the polarized light is irradiated in the opposite direction; y When polarized light is incident from the front, it first passes through the rectangular silicon pillar structure array to form a right-handed circularly polarized diffraction spot, and then passes through the metal grating structure. y Polarized light is transmitted, thus forming a y Polarized diffraction spot; when y When polarized light is incident from the back, it first passes through the metal grating structure. y Polarized light can pass through the metal grating structure and then pass through the rectangular silicon column structure array to form a left-handed circularly polarized diffraction spot on the detection plane. y Bidirectional image recognition is achieved under polarization state, thus realizing polarization-switchable image recognition function.

2. The polarization-switchable metasurface device with image recognition function according to claim 1, characterized in that: The rectangular silicon pillar structures in the rectangular silicon pillar structure array are designed according to transmittance requirements and can operate at different light wavelengths.

3. The metasurface device with polarization-switchable image recognition function according to claim 1, characterized in that: The metal digital imaging plate is made by coating a layer of gold film on a silicon dioxide substrate using vacuum coating technology. The thickness of the silicon dioxide substrate is 500 microns and the thickness of the gold film is 5 microns.

4. The polarization-switchable metasurface device with image recognition function according to claim 1, characterized in that: The dielectric substrate is made of silicon material with a thickness of 500 microns and a structure length and width of 13,000 microns.

5. The polarization-switchable metasurface device with image recognition function according to claim 1, characterized in that: Incident light use x Direction and y Two linearly polarized lights.

6. A method for realizing polarization-switchable image recognition based on a diffractive neural network, characterized in that: The implementation of a metasurface device with a polarization-switchable image recognition function according to any one of claims 1 to 5 comprises the following steps: The diffraction neural network consists of three parts: input layer, diffraction layer, and output layer. The gradient descent algorithm and back propagation algorithm are used to train and optimize the diffraction layer parameters. The recognition task data set is used to train the diffraction neural network diffraction layer parameters and determine the diffraction layer phase distribution; the optimized diffraction layer phase distribution text file is extracted, and the rectangular silicon pillar structure array is constructed by combining the dynamic phase with the geometric phase principle. The rectangular silicon pillar structure in the designed rectangular silicon pillar structure array realizes the function of a quarter wave plate at a specific frequency point, that is, the phase difference of the light wave in the fast axis and the slow axis direction is 90 degrees. This phase difference will change the polarization state of the light. Specifically, when linearly polarized light is incident on the optical axis of the quarter wave plate at an angle of 45 degrees, the light wave is decomposed into components along the fast axis and the slow axis. These two components will have 90 degrees after passing through the wave plate. The phase difference of degrees is combined into circularly polarized light; the polarization conversion function is realized by rotating each rectangular silicon pillar structure in the rectangular silicon pillar structure array by 45 degrees; thereby generating a metasurface device structure file; at the same time, the pattern to be recognized is processed, the pattern is hollowed out with a gold film, a silicon dioxide substrate is added, and the distance from the metasurface layer is set, and the substrate is placed parallel to the metasurface layer. The irradiation light source passes through the metal imaging plate and the metasurface layer in turn, and the metasurface device design with polarization switchable image recognition function is realized; The input layer of the diffraction neural network encodes the image information to be recognized and converts it into optical signals; these optical signals carry the amplitude distribution information of the image shape and are input into the next diffraction layer for processing; the optical signal is transmitted from the input layer to each diffraction layer, and the layers are interconnected through neurons to simulate the amplitude of the modulated light beam; in the diffraction layer, the optical signal interacts with these layers layer by layer and is processed through a multi-layer diffraction structure; each diffraction layer is modulated according to the optical signal transmitted by the previous layer to extract and enhance the characteristic information of the image; the output layer is responsible for detecting the light waves processed by all diffraction layers and performing image recognition tasks; in order to achieve the best target recognition effect, the diffraction layer obtains the optimal diffraction layer phase parameters through training; these parameters are applied to the distribution of the rectangular silicon pillar structure array to form a metasurface layer with polarization switchable function; the metasurface device can ultimately achieve efficient image recognition function under different polarization states, thereby improving the accuracy and efficiency of image recognition; The specific training process of the diffractive neural network model is as follows: Step 1) In the diffraction neural network model, the phase and amplitude changes of the light wave are calculated according to the Rayleigh-Sommerfeld diffraction integral formula. At the same time, the network model determines the bias term through the transmission coefficient of each metasurface layer. The model uses the forward propagation process to guide the incident light wave carrying the target information through a multi-layer metasurface structure with a specific configuration to form a diffraction pattern reflecting the characteristics of the target object on the detection plane. The Rayleigh-Sommerfeld diffraction theory describes the diffraction of the light field from the first n Layer to n +1 layer of propagation, where t ( r n )= Indicates n The transmission coefficient of the layer, where A is the amplitude, is the phase, i is an imaginary unit. When only the pure phase-type diffraction neural network is trained, = 1; the light field distribution in the propagation form of Rayleigh-Sommerfeld diffraction theory is expressed as: （1） in, U ( r n ) is the n The transmission electric field distribution of the layer, U ( r n+1 ) is the n Transmitted electric field distribution on the +1 layer, n is the number of diffraction layers, and h ( r n+1 -r n ) represents the impulse response during the transmission process, and the expression is: （2） in, It is from n Points on layer r n To n +1 layer of points r n+1 The Euclidean distance between λ is the incident wavelength, i = is an imaginary unit, x n+1 , y n+1 , z n+1 At a certain point in n +1 layer spatial coordinate components, , , They are n Layer The spatial coordinate components of the neurons; S represents the integration area, which is all the points within the entire wavefront area; the integral represents the integration of the entire n The light field distribution of the first layer is weighted and superimposed to calculate the n +1 layer of light field distribution at a certain point; this method can be used to predict and calculate the propagation of light waves in media with different layer structures; the number of diffraction layers is used n Indicates that a single diffraction layer is used here n = 1, use the n +1 indicates the detection plane, according to the energy distribution of the sub-detection area of ​​the detection plane S j = | U j | 2 , j The sub-detector regions are numbered, and the recognition result is determined by the category corresponding to the region with the highest energy; Step 2) Select a commonly used image classification dataset as the training dataset, and select the images to be identified and classified as the training set. The purpose is to train the diffraction layer phase parameters of the diffraction neural network to maximize the energy distribution of the corresponding category sub-detection area and minimize the energy distribution of other sub-areas; in the optimization process, the mean square error loss function is used MSE To evaluate the difference between the energy distribution of different sub-detector areas and the target energy distribution, the number of targets is identified by C express: （3） in s c represents the energy distribution received in the final detection area, represents the target energy distribution, represents the error between the target area and the detection area; based on the obtained diffraction pattern and its deviation from the expected result, a loss function is defined to evaluate the recognition accuracy of the diffraction neural network under the current parameters; Step 3) Apply the back propagation module and combine it with the gradient descent algorithm to calculate the gradient of the loss function with respect to the phase parameter of the diffraction layer to guide the optimization process of the phase distribution; after multiple iterations until Stop training when convergence occurs and the accuracy cannot be improved; repeat the above steps to adjust the number of diffraction layers to find the number of diffraction layers that can achieve the object recognition accuracy required by the designer; Step 4) iteratively executing step 1) to step 3), continuously updating the diffraction layer phase parameter until a predetermined recognition accuracy is met, and completing the training and optimization of the diffraction neural network; According to the phase distribution of the diffraction layer determined in step 2), eight rectangular silicon pillar structures of different sizes in the rectangular silicon pillar structure array are selected, which have a quarter wave plate function at a set frequency point, and rotating them 45 degrees can realize the function of polarization conversion, and they are arranged in an array according to the calculated phase distribution to form a multi-layer metasurface layer; the unit structure parameters of the supersurface layers include the length L, width W, height H and period length P of the rectangular silicon pillar structures in the rectangular silicon pillar structure array; the performance of rectangular silicon pillar structures of different sizes under linear polarized light irradiation is calculated to select the optimal parameter combination required at a specific wavelength and determine the size of the selected rectangular silicon pillar structure; according to the geometric phase principle, the expected phase distribution is achieved by rotating each rectangular silicon pillar structure; finally, the rectangular silicon pillar structure array is arranged according to the phase distribution obtained by training the diffraction neural network diffraction layer, and a metasurface device with polarization switchable image recognition function is constructed based on this; The final phase distribution is ,in is the phase obtained by training the diffraction neural network.

7. The method for realizing polarization-switchable image recognition based on diffractive neural network according to claim 6, characterized in that: The diffraction neural network training process uses a single diffraction layer, sets the distance from the input layer to the diffraction layer to 10,000 microns, the distance from the diffraction layer to the output layer to 10,000 microns, the learning rate to 0.001, the size of the neuron to 130 microns, and in order to save training time, eight samples are taken as a batch and input into the diffraction neural network.

8. The method for realizing polarization-switchable image recognition based on diffractive neural network according to claim 6, characterized in that: The training model software uses the TensorFlow framework, which is based on the Python programming language.

9. The method for realizing polarization-switchable image recognition based on diffractive neural network according to claim 6, characterized in that: Electromagnetic full-wave simulation software can be used to calculate the performance of rectangular silicon pillar structures in arrays of rectangular silicon pillar structures of different sizes under linear polarized light illumination.

Images