An all-optical image recognition detector and method based on a folded super-structured surface

Abstract

The method discloses a kind of full light image recognition detector and method based on folding superstructure surface, by multiple layers folding superstructure surface is stacked after by diffraction and reflection mode, and is integrated with CMOS image processing chip;The folding superstructure surface and CMOS image processing chip between chip are adhered by optical cement.The folding superstructure surface of each layer is composed of transparent dielectric substrate, silver reflection layer, superstructure unit array, the silver reflection layer is arranged on transparent dielectric substrate, and superstructure unit array is arranged with array on silver reflection layer;Each superstructure unit array is connected by light diffraction, light reflection, each layer folding superstructure surface is connected based on light diffraction, and the interconnection between each layer folding superstructure surface is realized using light diffraction, light reflection;The superstructure unit array is composed of dielectric nanometer column structure unit array arrangement, and the robustness of the neural network and the accuracy of image recognition can be greatly improved by the present application.

Description

TECHNICAL FIELD

[0001] The method belongs to the application technical field of micro-nano optics, and relates to a full-optical image recognition detector and method based on a folded superstructure surface. BACKGROUND

[0002] At present, scholars at home and abroad have very outstanding work on image processing of multi-layer diffraction deep neural network, which can simulate full connection neural network, convolutional neural network and the like in traditional deep learning, and the effect is pleasing. However, the current research cannot realize the chipization of the diffraction deep neural network. In addition, most scholars have proposed prospects for using superstructure surfaces as the carrier of photonic neural networks, but they have not really played the advantages of superstructure surfaces.

[0003] In order to improve the recognition accuracy of the diffraction deep neural network detector, increasing the number of diffraction layers is a relatively mainstream method. At present, the physical architecture of the multi-layer diffraction deep neural network is built in a vertical stacking manner, which will make the entire network become bloated and difficult to adapt to the working conditions with high volume requirements, and also becomes a big challenge for the chipization of the diffraction deep neural network. In recent years, due to the rise of artificial intelligence combined with optics and the emergence of superstructure materials, the superstructure surface as a two-dimensional form of superstructure materials can realize almost arbitrary electromagnetic wave front control, which has a natural degree of fit with the diffraction neural network. SUMMARY

[0004] Based on the technical ability of the current superstructure surface to completely independently control the light wave front and the technical characteristics of ultra-thin and ultra-flat, a folded superstructure surface scheme is designed, which provides a feasible scheme for realizing the chipization of the diffraction deep neural network and promotes the development of photonic computing.

[0005] In order to achieve the above purpose, the technical scheme adopted by the present application is a full-optical image recognition detector based on a folded superstructure surface, which is stacked by multiple folded superstructure surfaces through diffraction and reflection modes and integrated with a CMOS image processing chip.

[0006] Each folded superstructure surface is composed of a transparent dielectric substrate, a silver reflective layer and a superunit array, the silver reflective layer is arranged on the transparent dielectric substrate, and the superunit array is arranged on the silver reflective layer in an array; the superunit arrays are connected through light diffraction and light reflection, and the folded superstructure surfaces are connected based on light diffraction (two-dimensional), and the light diffraction and light reflection are used to realize the interconnection between the folded superstructure surfaces; the superunit array is composed of a dielectric nanocolumn structure unit array.

[0007] Furthermore, based on optical diffraction and reflection design, the physical architecture of the deep diffraction neural network of the folded metasurface is obtained, and the parameters of each neuron in this physical architecture are determined. A mathematical model of the physical architecture is constructed using deep learning algorithms, and the neuron parameters are calculated, optimized, iterated, and updated to obtain the phase information of each dielectric nanopillar structural unit in each layer of the folded metasurface used for image recognition, thus obtaining the final neuron parameters of the deep diffraction neural network. The neuron parameters refer to optical phase and amplitude information, corresponding to the optical phase and amplitude information between the dielectric nanopillar structural unit arrays in the folded metasurface.

[0008] Furthermore, in the optical diffraction connection of each layer of the folded metasurface, the optical phase of the neurons is used as the weight, and each point in the physical model of the folded metasurface is a sub-wave source of a secondary wave. The input of a neuron in the next layer of the folded metasurface is defined as the superposition result of the outputs of all neurons in the previous layer after diffraction in free space at that neuron. A pure phase-controlled neural network is used. The neuron refers to a dielectric nanopillar structure unit.

[0009] Furthermore, when integrating the folded metasurfaces of each layer, the distance R between the folded metasurfaces of each layer is determined according to the training process of the deep diffraction neural network, i.e. Where d is the distance between the two reflective layers. The deflection angle is the angle of the signal light after being modulated by the dielectric nanopillar structure of the folded metasurface.

[0010] Furthermore, the phase values ​​of neurons are trained using a diffraction propagation function, i.e., the output of the network is calculated through optical diffraction forward propagation, and the objective function is optimized by combining the error backpropagation algorithm to realize a deep diffraction neural network for the folded metasurface. The optical phase information of neurons in each layer of the folded metasurface is obtained, and after comparison and selection with the database, a dielectric nanopillar structure is obtained.

[0011] Furthermore, the signal light is a single wavelength of 532nm visible light used for image recognition, and is realized using a reflective metasurface composed of the folded meta-unit array connected by light refraction. Considering that the signal light is incident at an angle onto the folded metasurface, the reflective metasurface uses a PB phase form to adjust the phase information of the incident light in order to improve diffraction efficiency.

[0012] Further, in the database, the dielectric nanocolumn structure is simulated by using computer FDTD software, the optical phase and amplitude information change of the dielectric nanocolumn structure generated by the incident light of 532 nm wavelength are obtained, and then a group of dielectric nanocolumn structures with high conversion efficiency are selected as candidate structures in the database. By adjusting the incident light tilt angle, the optical phase difference of each candidate structure is obtained by subtracting the basic phase generated by each candidate structure from the generated optical phase. According to the comprehensive consideration of the basic phase and the phase difference, the structure and size with the smallest actual phase difference sum are selected as the final size of the folded superstructure surface. The structure and layout of each folded superstructure surface are designed to prepare a sample, and a conformal filling process for high-precision small-area processing is adopted to manufacture a high-precision superstructure surface through electron beam exposure (EBL) and atomic layer deposition.

[0013] Further, the folded superstructure surface is prepared by using TiO2 which has excellent optical response in the corresponding waveband.

[0014] Based on the advantages of the super-thin and super-flat of the superstructure surface, the optical reflection and refraction are used to realize the optical interconnection between the diffraction layers, that is, the folded superstructure surface is used to carry the diffraction deep neural network detector, which greatly reduces the volume of the device. Furthermore, it is integrated with the CMOS image processing chip, and the diffraction deep neural network is chipped, which is undoubtedly an excellent solution to the above problems.

[0015] This technology realizes the information transmission between the multi-layer superstructure surfaces based on the optical reflection principle, and is integrated with the CMOS photosensitive device, wherein each layer of the superstructure surface is composed of a transparent dielectric substrate, a silver reflective layer and a two-dimensional array of dielectric nanocolumn structures. A full-optical deep diffraction neural network satisfying a specific function is trained by a deep diffraction neural network algorithm, and the physical information of the entire superstructure surface is calculated. Then, each superstructure surface is processed and manufactured according to the design requirements. Finally, the multi-layer superstructure surface and the CMOS photosensitive chip are integrated, and the final target recognition task is realized after being processed by the COMS processor.

[0016] Compared with the existing optical deep diffraction neural network, the present application greatly reduces the physical size of the neural network and provides higher experimental value for realizing the chip of the optical deep diffraction neural network.

[0017] The purpose of the present application is to provide a full-optical image recognition detector based on a folded superstructure surface, which realizes the optical interconnection between the superstructure surfaces by using reflection and refraction, and is integrated with the CMOS image processing chip to realize image recognition.

[0018] The present application has the following advantages and effects compared with the existing technology:

[0019] 1、The present application adopts a multi-layer diffraction neural network to train the phase of the super-structured surface, and can greatly improve the robustness of the neural network and the accuracy of image recognition through effective physical means, so that it can be comparable to the longitudinal stacked diffraction deep neural network.

[0020] 2、The present application innovatively uses a folding physical architecture to model a multi-layer diffraction neural network, greatly reducing the size of the device itself under the premise of ensuring high-precision recognition accuracy.

[0021] 3、The present application directly integrates a multi-layer super-structured surface on a CMOS image processing chip, which can directly extract light intensity information, providing a more feasible solution for realizing a diffraction deep neural network chip. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 The design flowchart and preparation flowchart of the present application.

[0023] Figure 2 The device principle diagram of the present application, wherein the super-structured surface 1 is only used to realize beam deflection; the super-structured surfaces 2 and 3 are obtained through deep learning training, and play a role in modulating the amplitude and / or phase of the signal light; and the device 4 is a CMOS image processing chip light-sensitive area.

[0024] Figure 3 The structure schematic diagram of the super-structured surface unit in the present application.

[0025] Figure 4 The schematic diagram of the tilt angle spectrum transmission mode based on angular spectrum diffraction transmission.

[0026] Figure 5 The overall schematic diagram of the algorithm design, processing integration and evaluation test of the present application. DETAILED DESCRIPTION

[0027] The present application will be described in detail below in combination with the drawings and examples.

[0028] Further, the implementation method of the folded super-structured surface all-optical image recognition detector includes a folded super-structured surface design method and a folded super-structured surface preparation method.

[0029] The folding super-structure design method comprises the following steps: designing a physical architecture of a deep diffraction neural network of the folding super-structure based on optical reflection and refraction, and determining parameters of each neuron of the physical architecture; constructing a physical model of the folding super-structure by using a deep learning algorithm, calculating, optimizing, iterating and updating the parameters of the neurons, obtaining phase information of each unit in each layer of the folding super-structure for image recognition, and obtaining final parameters of the deep diffraction neural network. The parameters of the neurons refer to optical phase and amplitude information corresponding to the optical phase and amplitude information between the unit arrays of the dielectric nanocolumn structure of the folding super-structure.

[0030] The phase of the neuron is used as the weight by using diffraction of light passing through the diffraction layer to connect each level, each point on the diffraction layer is a sub-wave source of a secondary wave, and the input of a neuron in the next layer is defined as the superposition result of the outputs of all neurons in the previous layer after free-space diffraction on the neuron. A pure phase control type neural network is adopted.

[0031] The phase value of the neuron is trained by using a specific diffraction propagation function, that is, the output of the network is calculated by optical diffraction, an error back propagation algorithm is combined to optimize an objective function, an ideal network architecture is designed, the phase information of each unit of the super-structure in each layer is obtained, a nanocolumn structure with a specified phase is matched by searching a database, and the distance between layers is strictly in accordance with the training process of the diffraction neural network when the super-structure is integrated. wherein d is the distance between two reflection layers, is a deflection angle of signal light after being modulated by the super-structure 1.

[0032] In the folding super-structure design method, the signal light for image recognition is a single wavelength of 532nm visible light, and the reflection type super-structure composed of the super-structure unit arrays connected by light refraction is adopted. Considering that the signal light is obliquely incident on the folding super-structure, the reflection type super-structure adopts a P-B phase form to adjust the phase information of the incident light, so as to improve the diffraction efficiency.

[0033] The folding super-structure design method comprises the following steps: according to the principle of Mie resonance, a dielectric nano-pillar structure is simulated by using a computer FDTD software, and a database is established based on the optical phase and amplitude information changes of the dielectric nano-pillar structure generated by incident light with a wavelength of 532 nm. A plurality of groups of structures with high conversion efficiency are selected as candidate structures in the database, and the phase difference of each candidate structure is obtained by adjusting the tilt angle of the incident light and subtracting the basic phase generated by each candidate structure from the generated phase. According to the comprehensive consideration of the basic phase and the phase difference, the structure and size with the smallest actual phase difference sum are selected as the final size of the folding super-structure. The structure and layout of each folding super-structure are designed to prepare a sample, a conformal filling process for high-precision small-area processing is adopted, and a high-precision super-structure is manufactured by electron beam exposure (EBL) and atomic layer deposition.

[0034] The folding super-structure preparation method comprises the following steps: a material (TiO2) with excellent optical response in the corresponding waveband is used to realize the preparation of the super-structure. The corresponding layout of each folding super-structure to be prepared is drawn, the required pattern is processed by electron beam exposure, then TiO2 is wrapped in the resist structure by atomic layer deposition, finally the surface is filled by using a sacrificial layer, the surface is polished to a specified height by ion beam etching, and then the resist is removed to realize pattern transfer. The prepared folding super-structure is adhered to a CMOS image processing chip by using a special optical adhesive, and a detector for image recognition is obtained.

[0035] The folding super-structure preparation method comprises the following steps: the super-structure of the present application only uses visible light with a wavelength of 532 nm as experimental signal light, but can be used in a large waveband range from ultraviolet to infrared. Different refractive indexes are constructed by selecting nano structures with different sizes, which are matched with the database, and the image recognition has the advantages of wideband.

[0036] Embodiment

[0037] The implementation method of the full-optical image recognition detector based on the folding super-structure mainly comprises three parts: design of the folding super-structure, design and manufacture of the super-structure, and integration scheme with a CMOS image processing chip, and the specific process is as shown in Figure 1 .

[0038] The design of the folding super-structure and the implementation method of image recognition: the physical architecture of the folding super-structure deep diffraction neural network is designed by using the principles of optical diffraction and reflection, as shown in Figure 2As shown, the parameters of the physical architecture are determined; a physical model of the folded metasurface is constructed using a deep learning algorithm, the neuron parameters are calculated, optimized, iterated and updated, the phase information of each unit in each layer of the metasurface for image recognition is obtained, and the final parameters of the neural network are obtained. The light signal used for image recognition this time is a single wavelength of 532nm visible light, a reflective metasurface is used, and the signal light of this scheme is incident on the metasurface at an angle, so the P-B phase form of the metasurface unit is used to adjust the phase information of the incident light to improve the diffraction efficiency. According to the principle of Mie resonance, the FDTD software is used to simulate the phase change of the elliptical nanocolumn with a certain height and the size of the long and short axes in the range of 30nm to 150nm under the incident light of 532nm wavelength, thereby establishing a database. A plurality of groups of structures with high polarization conversion efficiency are selected from the database as candidate structures, the incident light is adjusted by a small angle, the phase difference between the generated phase and the basic phase of each candidate structure is obtained, and the final structure size is selected according to the comprehensive consideration of the basic phase and the phase difference. The structure and layout of each diffraction layer are designed to prepare the sample, and the conformal filling process for high-precision small-area processing is adopted to manufacture the high-precision metasurface through electron beam exposure (EBL) and atomic layer deposition.

[0039] A method for manufacturing a folded metasurface-based all-optical image recognition detector: a material (TiO2) with excellent optical response in the corresponding waveband is used to prepare the metasurface. The layout of each folded metasurface to be prepared is drawn, the required pattern is processed by electron beam exposure, then TiO2 is wrapped in the resist structure by atomic layer deposition, finally the surface is filled flat by using a sacrificial layer, and then the surface is polished to a specified height by ion beam etching, and then the resist is removed to realize pattern transfer. The prepared side faces the CMOS image processing chip, and the metasurface is adhered to the CMOS image processing chip by using a special optical adhesive, thereby obtaining a detector for image recognition.

[0040] In order to make the principles and advantages of the present application clearer and more apparent, the present application is further described in detail by taking the design of handwritten digit 0 and handwritten digit 1 as image recognition objects as an example. It should be understood that the description herein is only used to explain the present application and is not used to limit the present application.

[0041] Figure 1 The design flowchart and preparation flowchart of the present application.

[0042] Figure 2This is a schematic diagram of the device of the present invention. Metasurface 1 is used only to deflect the light beam; metasurfaces 2 and 3 are obtained through deep learning training and function to modulate the amplitude and / or phase of the signal light; device 4 is the photosensitive area of ​​the CMOS image processing chip. The present invention uses a combination of diffraction and reflection to achieve interconnection between diffraction layers, thereby greatly reducing the size of the diffraction neural network.

[0043] Figure 3 This is a schematic diagram of the structure of the metasurface unit in this invention.

[0044] The forward propagation process of the diffraction neural network is based on the angular spectrum diffraction theory in Fourier optics. First, the impulse response function, i.e., the transfer function, of the Rayleigh-Sommerfeld diffraction formula is given as the forward propagation calculation formula:

[0045]

[0046] in This is the diffraction distance.

[0047] Considering that the signal light between the various diffraction layers in this patent is not perpendicular to the diffraction surface, the classic angular spectrum diffraction theory cannot be used. Therefore, this invention derives a tilted angular spectrum transmission method suitable for this patent based on angular spectrum diffraction transmission. For example... Figure 4 As shown, surface P1 is the light source surface, and surface P2 is the intermediate conversion surface perpendicular to the optical axis. After the frequency domain signal of the signal light on surface P1 is converted to that on surface P2, the light field can be diffracted and transmitted between P2 and P3 using the traditional diffraction transmission method between parallel planes.

[0048] On surface P1, the complex amplitude of light in the frequency domain is G(f x ,f y If ,0), then its inverse transform spatial distribution is:

[0049]

[0050] The above equation shows that g(x,y,0) can be viewed as a superposition of a series of wavefronts:

[0051] u(x,y,z=0;f x ,f y )=G(f x ,f y ,0)exp[i2π(f x ·x+f y ·y)] (3)

[0052] Among them, G(f x ,f y ,0) can represent the complex amplitude of a plane wave.

[0053] A plane wave with amplitude A and wave vector k can be expressed as:

[0054]

[0055]

[0056] where: Combining equations (3) and (4) gives:

[0057] k = 2π[f x ,f y ,w(f x ,f y )] (6)

[0058] where,

[0059] Similarly, the spectrum of the complex amplitude distribution on the P2 plane can be expressed as:

[0060]

[0061] Similarly, equation 6 can be obtained as:

[0062]

[0063] where,

[0064] The wave vector k in equation 6, 8, is a vector in the coordinate system (x, y, z) and respectively, and these two wave vectors can be transformed into each other through coordinate transformation. Assuming that the matrix T is the transformation matrix between the P1 and P2 planes, the transformation between these two wave vectors is:

[0065]

[0066] Assuming the transformation matrix is:

[0067]

[0068] This means that the transformation relationship between the Fourier frequency domain parameters of the P1 and P2 planes is:

[0069]

[0070]

[0071] Therefore, the spectrum on the P1 plane coordinate system can be expressed using the spectrum on the P2 plane coordinate system:

[0072]

[0073] right The light intensity on the P2 surface can be obtained by performing an inverse Fourier transform. However, after frequency space coordinate transformation, the inverse Fourier transform is used to represent... When dealing with a light wave field, according to the substitution method of double integrals, the integral surface must satisfy... The mathematical relationship, in which For the Jacobian determinant:

[0074]

[0075] Therefore, when using coordinate transformation in frequency space to represent When the light wave field is on the surface, its expression is:

[0076]

[0077] Since the beam deflection discussed in this paper involves the P2 plane rotating counterclockwise around the X-axis by an angle... Therefore, the transformation matrix is:

[0078]

[0079] Substituting equation 16 into equations 11 and 12, we get:

[0080]

[0081]

[0082] Then the angular spectrum and Jacobi determinant on plane P2 are respectively:

[0083]

[0084]

[0085] Therefore, the complex amplitude on the P2 plane is:

[0086]

[0087] The above demonstrates the transformation mechanism of the signal light through tilted diffraction after reflection. The subsequent connections between the various diffraction layers remain consistent with classical diffraction theory.

[0088] The complex amplitude function of the light field in the (l+1)th layer of a diffraction neural network is defined as follows:

[0089]

[0090] in For the incident light field of the l-th layer, For the Jones matrix of unit structure, it can be derived that:

[0091]

[0092] Further, the light intensity distribution of the output light field of the last layer of diffraction layers is obtained:

[0093]

[0094] After the forward propagation, the reverse optimization process is performed, the loss function is defined as the mean square error of the last layer output light field and the theoretical target light field, the output light field is gradientized and simplified to obtain the Lth layer:

[0095]

[0096] As described above, the phase value of the neuron is trained by using a specific diffraction propagation function, that is, the output of the network is calculated by optical diffraction in the forward direction, and the error back propagation algorithm is used to optimize the objective function, the design of the ideal network architecture is realized, the phase information of each layer of super-structured surface unit is obtained, and the nano pillar structure with specified phase is matched by searching the established database, that is Figure 3 The structure parameters a, b and theta. At this time, a diffraction neural network based on super-structured surface is successfully built, and a folding diffraction neural network with image recognition is obtained as shown in Figure 2 When the super-structured surface is integrated, the distance between layers is strictly in accordance with the process of training the diffraction neural network, that is Wherein, d is the distance between two reflection layers, Is the deflection angle of the signal light after being modulated by the super-structured surface 1.

[0097] Figure 5 It is the research content system framework of the present application, when designing the diffraction neural network, the present application analogs the implementation process of the traditional neural network, based on the Huygens-Fresnel theory, uses the diffraction of light through the diffraction layer to connect each level, takes the phase of the neuron as the weight, and each point on the diffraction layer is a sub-wave source of a secondary wave, the input of a neuron in the next layer is defined as the superposition result of the output of all neurons in the last layer after free space diffraction at the neuron, and the present application adopts a pure phase control type neural network.

[0098] The above introduces the design, manufacturing and functional principle of the super structured surface, and the following is the on-chip integration method: according to the different working distances of the detector, when the working distance is small, the optical transparent adhesive can be directly used for gluing with the CMOS image processor to keep a short distance, when the working distance is large, the optical glass with a specific thickness is customized, and the above gluing process is repeated to realize the integration, and the detector based on the multi-layer super structured surface diffraction for target identification and tracking is obtained. In the full-optical image recognition method based on the folding super structured surface, the optical diffraction and reflection are used to realize the optical interconnection of the plane transverse distribution super structured surface and the integration with the CMOS image processing chip, the recognition of the target image can be realized, and the method has the advantages of fast response, ultra-low power consumption, compact structure, high integration and the like.

Claims

1. A plenoptic image recognition detector based on folded metasurface, characterized in that, The multi-layer folded super-structure is stacked by diffraction and reflection, and is integrated with a CMOS image processing chip; the folded super-structure and the CMOS image processing chip are adhered by optical glue; Each layer of the folded super-structure is composed of a transparent dielectric substrate, a silver reflective layer and an array of super-structure units; the silver reflective layer is arranged on the transparent dielectric substrate, and the array of super-structure units is arranged on the silver reflective layer; the arrays of super-structure units are connected by light diffraction and light reflection; the layers of the folded super-structure are connected by light diffraction; the interconnection between the layers of the folded super-structure is realized by light diffraction and light reflection; the array of super-structure units is composed of an array of dielectric nano-pillar structure units; The folded super-structure is designed based on light diffraction and reflection; the physical architecture of the deep diffraction neural network of the folded super-structure is obtained, and the parameters of each neuron of the physical architecture are determined; a mathematical model of the physical architecture is constructed by using a deep learning algorithm; the parameters of the neurons are calculated, optimized, iterated and updated; the phase information of each dielectric nano-pillar structure unit in each layer of the folded super-structure for image recognition is obtained; the final parameters of the neurons of the deep diffraction neural network are obtained; the parameters of the neurons refer to optical phase and amplitude information, which correspond to the optical phase and amplitude information between the arrays of dielectric nano-pillar structure units in the folded super-structure; In the light diffraction connection between the layers of the folded super-structure, the optical phase of the neuron is used as the weight; each point on the physical model of the folded super-structure is a sub-wave source of a secondary wave; the input of a neuron of the next layer of the folded super-structure is defined as the superposition result of the outputs of all neurons of the previous layer after free-space diffraction; a pure phase control type neural network is used. In the integration of each layer of the folded metasurface, the folded metasurface distance R between layers is in accordance with the training process of the depth diffraction neural network, that is wherein d is the distance between two reflection layers, is the deflection angle of the signal light modulated by the dielectric nanocolumn structure of the folded metasurface. 2.The all-optical image recognition detector based on the folded metasurface of claim 1, wherein The phase value of the neuron is trained by using a diffraction propagation function; the output of the network is calculated by using optical diffraction; an error back propagation algorithm is used to optimize the objective function; the deep diffraction neural network of the folded super-structure is realized; the optical phase information of the neurons of each layer of the folded super-structure is obtained; the dielectric nano-pillar structure is obtained after comparison and selection in the database. 3.The all-optical image recognition detector based on the folded metasurface of claim 1, wherein, The signal light is a single wavelength of 532nm visible light for image recognition; a reflective super-structure composed of the light refraction connection between the arrays of super-structure units is used to realize the signal light; considering that the signal light is obliquely incident on the folded super-structure, the P-B phase form is used to adjust the phase information of the incident light in the reflective super-structure to improve the diffraction efficiency.

4. The all-optical image recognition detector based on the folded metasurface according to claim 2, characterized in that, In the database, the dielectric nano-pillar structure is simulated by using a computer FDTD software; the change of the optical phase and amplitude information of the dielectric nano-pillar structure generated by the incident light of 532nm wavelength is used to complete the establishment; In the database, a plurality of groups of dielectric nano-pillar structures with high conversion efficiency are selected as candidate structures; the optical phase generated by adjusting the incident light oblique angle is subtracted from the basic phase generated by each candidate structure to obtain the phase difference of each candidate structure. According to the comprehensive consideration of the base phase and the phase difference, a structure and a size with the minimum sum of actual phase differences are selected as the final size of the folded super-structured surface; a structure and a layout of each folded super-structured surface are designed to prepare a sample; and a high-precision super-structured surface is manufactured through electron beam exposure and atomic layer deposition by using a conformal filling process. 5.The all-optical image recognition detector based on folded metasurface of claim 1, wherein, The folded super-structured surface adopts TiO2.

Images