Tunable metasurface structure, imager and target recognition method

Abstract

The present invention relates to the technical field of optical devices, and provides a tunable metasurface structure, an imager, and a target recognition method. The tunable metasurface structure includes a first electrode layer, a second electrode layer, a liquid crystal layer, and a metasurface layer; by applying voltage to the first electrode layer and the second electrode layer, the liquid crystal molecules in the liquid crystal layer are controlled to turn to adjust the polarization state of light in the incident image; the metasurface layer responds differently to light of different polarization states through polarization multiplexing, and the left-handed circular polarization state and the right-handed circular polarization state are used to encode the positive and negative values ​​of the weight, and a multi-channel A×A spot matrix is ​​generated through spatial multiplexing to determine the emitted characteristic image through the spot matrix; different spot brightness in the spot matrix is ​​used to characterize different convolution kernel weights. The present invention combines a metasurface and a liquid crystal to obtain a tunable structure, which can realize the processing of various coded information, thereby improving the accuracy and robustness of object recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of optical devices, and in particular to a tunable metasurface structure, an imager and a target recognition method. Background Art

[0002] The rise of Convolutional Neural Networks (CNNs) and the emergence of large-scale training datasets have laid a solid foundation for many artificial intelligence applications. Although these advances have significantly improved performance, they usually come at the cost of increasing model complexity, which in turn requires more computing resources.

[0003] As a result, the growth in the use and complexity of neural networks has led to an increase in energy consumption, especially when computing power is limited, which poses challenges for real-time decision-making. For example, for machine vision, processors and imagers need to be compact, lightweight and energy-efficient to support onboard processing while ensuring low power consumption and high throughput parallel operation. Therefore, these conflicting requirements highlight the need to develop innovative software and hardware solutions in the context of the ever-expanding needs of artificial intelligence.

[0004] Optical neural networks can increase computational speed and improve energy efficiency in a variety of applications. Machine vision computation can be significantly accelerated by integrating computational tasks into front-end imaging optics. Free-space optical computation, especially Fourier optics, predates modern digital circuits and enables highly parallel convolution operations common in machine vision architectures.

[0005] Optical neural networks have promoted breakthrough advances in machine vision, smart healthcare, and autonomous driving systems. However, fixed optical network architectures have encountered great challenges in achieving large-scale parameter adjustments without modifying physical components, limiting their ability to handle complex application scenarios, and the accuracy and stability of optical neural networks are still lacking. Summary of the invention

[0006] The present invention provides a tunable metasurface structure, an imager and a target recognition method, which are used to solve the technical problems of low accuracy and stability in traditional optical neural networks in the prior art.

[0007] The present invention provides a tunable metasurface structure, comprising a first electrode layer, a second electrode layer, a liquid crystal layer and a metasurface layer; the first electrode layer is arranged on the metasurface layer; the liquid crystal layer is arranged on the first electrode layer; the second electrode layer is arranged on the liquid crystal layer; wherein, by applying voltage to the first electrode layer and the second electrode layer, the steering of liquid crystal molecules in the liquid crystal layer is controlled to adjust the polarization state of light in an incident image; the metasurface layer responds differently to light of different polarization states through polarization multiplexing, the left-handed circular polarization state and the right-handed circular polarization state are used to encode the positive and negative values ​​of the weight, and a multi-channel A×A spot matrix is ​​generated through spatial multiplexing to determine the emitted characteristic image through the spot matrix; wherein A represents the number of rows and columns of the spot matrix; different spot brightnesses in the spot matrix are used to characterize different convolution kernel weights; the tunable metasurface structure is expressed as:

[0008] ;

[0009] in, represents the operating wavelength; x and y represent the position information of the abscissa and ordinate in the plane of the metasurface layer respectively; represents the phase of the i-th spot in the spot matrix, represents the focal length of the metasurface layer, It represents the offset distance of the i-th spot in the spot matrix relative to the central spot in the x direction. Indicates the offset distance of the i-th spot in the spot matrix relative to the central spot in the y direction.

[0010] According to a tunable metasurface structure provided by the present invention, adjusting the polarization state of light in an incident image includes adjusting the ratio of the left-handed circular polarization state and the right-handed circular polarization state of the light in the incident image; in the optical neural network of the metasurface layer, the positive values ​​in the spot matrix correspond to the left-handed circular polarization state, and the negative values ​​in the spot matrix correspond to the right-handed circular polarization state.

[0011] According to a tunable metasurface structure provided by the present invention, the properties of the liquid crystal layer are expressed as follows:

[0012] ;in, Indicates the direction of rotation of liquid crystal molecules; represents the constant axis refractive index; represents the extraordinary refractive index; Represents the effective refractive index.

[0013] According to a tunable metasurface structure provided by the present invention, the circularly polarized light multiplexing of the metasurface layer is expressed as: ;in, represents the rotation angle of the nanopillars in the supersurface layer; Indicates when When set to 0, the phase delay experienced by the nanopillars along the x direction in the metasurface layer; represents the phase delay of the left circular polarization state; represents the phase delay of the right-hand circular polarization state.

[0014] A tunable metasurface structure provided according to the present invention also includes a first light-transmitting substrate layer and a second light-transmitting substrate layer; the metasurface layer is arranged on the first light-transmitting substrate layer; and the second light-transmitting substrate layer is arranged on the second electrode layer.

[0015] According to a tunable metasurface structure provided by the present invention, the metasurface layer is used to perform a dot product operation on an incident image and a numerical value represented in a light spot matrix to obtain a characteristic image.

[0016] According to a tunable metasurface structure provided by the present invention, the thickness of the liquid crystal layer ranges from 1 micron to 10 microns, the cross-sectional shape of the nanocolumn is rectangular or elliptical; the length of the cross-sectional shape of the nanocolumn ranges from 100 nanometers to 1000 nanometers; the height of the nanocolumn ranges from 100 nanometers to 1000 nanometers; and the distribution period of the nanocolumns ranges from 200 nanometers to 1000 nanometers.

[0017] According to a tunable metasurface structure provided by the present invention, materials of the first electrode layer and the second electrode layer are both indium tin oxide; materials of the first light-transmitting substrate layer and the second light-transmitting substrate layer are both glass.

[0018] The present invention also provides an imager based on an optical neural network, comprising a photodetector and the above-mentioned tunable metasurface structure; wherein the tunable metasurface structure is used to obtain a characteristic image after polarization multiplexing and spatial multiplexing of the light of the received image, and the photodetector is used to receive the light of the characteristic image of the tunable metasurface structure and form an image.

[0019] The present invention also provides a target recognition method based on an optoelectronic hybrid convolutional neural network, using the above-mentioned tunable metasurface structure. The target recognition method based on an optoelectronic hybrid convolutional neural network includes: performing convolution processing of an optical neural network on a target to be identified through a tunable metasurface structure to obtain a feature image after feature enhancement; converting the feature image from an optical signal to an electrical signal through a photoelectric detector; inputting the feature image of the electrical signal into a digital neural network for target recognition, and obtaining a recognition result output by the digital neural network.

[0020] The present invention provides a tunable metasurface structure, an imager and a target recognition method, wherein the tunable metasurface structure comprises a first electrode layer, a second electrode layer, a liquid crystal layer and a metasurface layer; the first electrode layer is arranged on the metasurface layer; the liquid crystal layer is arranged on the first electrode layer; the second electrode layer is arranged on the liquid crystal layer; by applying voltage to the first electrode layer and the second electrode layer, the liquid crystal molecules in the liquid crystal layer are controlled to turn to adjust the polarization state of the light in the incident image; the metasurface layer responds differently to light of different polarization states through polarization multiplexing, and the left-handed circular polarization state and the right-handed circular polarization state are used to encode the positive and negative values ​​of the weight, and a multi-channel A×A spot matrix is ​​generated through spatial multiplexing to determine the emitted characteristic image through the spot matrix; different spot brightness in the spot matrix is ​​used to characterize different convolution kernel weights. The present invention combines the metasurface and the liquid crystal to obtain a tunable structure, which can realize the processing of various coded information, thereby improving the accuracy and robustness of object recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0022] Figure 1 It is one of the structural schematic diagrams of the tunable metasurface structure provided in an embodiment of the present invention.

[0023] Figure 2 This is the second structural schematic diagram of the tunable metasurface structure provided in an embodiment of the present invention.

[0024] Figure 3 It is a schematic diagram of the structure of an imager based on an optical neural network provided in an embodiment of the present invention.

[0025] Figure 4 It is a schematic diagram of the structure of the super surface layer provided in an embodiment of the present invention.

[0026] Figure 5 It is a schematic diagram of the principle of an imager based on an optical neural network provided in an embodiment of the present invention.

[0027] Figure 6 It is a flow chart of a target recognition method based on an optoelectronic hybrid convolutional neural network provided in an embodiment of the present invention.

[0028] Figure 7 It is a structural schematic diagram of the optoelectronic hybrid convolutional neural network architecture provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] In the description of the embodiments of the present invention, it should be noted that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the embodiments of the present invention and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limitations on the embodiments of the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0031] In the description of the embodiments of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "connected" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the embodiments of the present invention can be understood according to specific circumstances.

[0032] In the embodiments of the present invention, unless otherwise clearly specified and limited, the first feature being "above" or "below" the second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, the first feature being "above", "above" and "above" the second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. The first feature being "below", "below" and "below" the second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0033] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiment of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0034] The present invention provides a tunable metasurface structure, see Figure 1 , Figure 1 1 is one of the structural schematic diagrams of the tunable metasurface structure provided by the embodiment of the present invention. In this embodiment, the tunable metasurface structure may include a first electrode layer 121 , a second electrode layer 122 , a liquid crystal layer 130 and a metasurface layer 110 .

[0035] Specifically, the first electrode layer 121 may be disposed on the super-surface layer 110 ; the liquid crystal layer 130 may be disposed on the first electrode layer 121 ; and the second electrode layer 122 may be disposed on the liquid crystal layer 130 .

[0036] By applying voltage to the first electrode layer 121 and the second electrode layer 122, the direction of the liquid crystal molecules in the liquid crystal layer 130 is controlled to adjust the polarization state of the light in the incident image; the metasurface layer 110 responds differently to light with different polarization states through polarization multiplexing, and the left-handed circular polarization state and the right-handed circular polarization state are used to encode the positive and negative values ​​of the weight, and a multi-channel A×A spot matrix is ​​generated through spatial multiplexing to determine the emitted characteristic image through the spot matrix.

[0037] Where A represents the number of rows and columns of the spot matrix; different spot brightness in the spot matrix is ​​used to represent different convolution kernel weights.

[0038] In the above manner, the embodiment of the present invention combines the metasurface and liquid crystal to obtain a tunable structure, which can process various coded information, thereby improving the accuracy and robustness of object recognition.

[0039] A metasurface is a nanoscale optical element that can integrate a wider range of wavelength properties, including phase, wavelength, angle, and amplitude, into optical computing. By integrating a polarization multiplexing metasurface with a photodetector, an optical convolution kernel can be generated. This embodiment integrates liquid crystal and metasurface into a tunable structure that can be adjusted by an applied voltage. By electrically controlling the orientation of the LC molecules, continuous intensity regulation of the point spread function (PSF) of the two spiral channels is achieved, thereby realizing multi-channel convolution in parallel through polarization multiplexing and spatial multiplexing.

[0040] Optionally, the thickness of the liquid crystal layer ranges from 1 micrometer to 10 micrometers.

[0041] Optionally, the cross-sectional shape of the nanocolumn is rectangular or elliptical; the length of the cross-sectional shape of the nanocolumn ranges from 100 nanometers to 1000 nanometers; the height of the nanocolumn ranges from 100 nanometers to 1000 nanometers; and the distribution period of the nanocolumn ranges from 200 nanometers to 1000 nanometers.

[0042] Furthermore, the tunable metasurface structure is expressed as:

[0043] ;

[0044] in, represents the operating wavelength; x and y represent the position information of the abscissa and ordinate in the plane of the metasurface layer respectively; represents the phase of the i-th spot in the spot matrix, represents the focal length of the metasurface layer, It represents the offset distance of the i-th spot in the spot matrix relative to the central spot in the x direction. Indicates the offset distance of the i-th spot in the spot matrix relative to the central spot in the y direction.

[0045] Optionally, the material of the first electrode layer 121 and the second electrode layer 122 may be a transparent conductive material, such as indium tin oxide (ITO), zinc oxide (ZnO) doped material, silver nanowires, and the like.

[0046] In some embodiments, the properties of the liquid crystal layer can be expressed as:

[0047] .

[0048] in, Indicates the direction of rotation of liquid crystal molecules; represents the constant axis refractive index; represents the extraordinary refractive index; Represents the effective refractive index.

[0049] In some embodiments, the circularly polarized light multiplexing of the metasurface layer can be expressed as:

[0050] .

[0051] in, represents the rotation angle of the nanopillars in the supersurface layer; Indicates when When set to 0, the phase delay experienced by the nanopillars along the x direction in the metasurface layer; represents the phase delay of the left circular polarization state; represents the phase delay of the right-hand circular polarization state.

[0052] In some embodiments, adjusting the polarization state of the light in the incident image includes adjusting the ratio of the left-circularly polarized light (LCP) and the right-circularly polarized light (RCP) of the light in the incident image.

[0053] Polarized light refers to light whose electric field vector of the light wave vibrates in a certain direction in the direction of propagation. Polarized light can be divided into linearly polarized light, circularly polarized light and elliptically polarized light. Among them, circularly polarized light can be divided into left-handed circularly polarized light and right-handed circularly polarized light. Left-handed circularly polarized light refers to light whose electric field vector rotates in a clockwise direction in the direction of propagation. Right-handed circularly polarized light refers to light whose electric field vector rotates in a counterclockwise direction in the direction of propagation.

[0054] In this embodiment, the arrangement direction of the liquid crystal molecules in the liquid crystal layer can be changed by controlling the voltage of the electrode layer, thereby controlling the polarization state of light.

[0055] For example, when the applied voltage is 0, the liquid crystal molecules are arranged along the initial direction, and the linearly polarized light remains in the linear polarization state after passing through the liquid crystal layer; when the voltage increases, the liquid crystal molecules gradually align along the direction of the electric field, and the light will be converted into left-handed circular polarization and right-handed circular polarization states in different proportions after passing through the liquid crystal layer.

[0056] Among them, in the optical neural network of the metasurface layer, left-handed circular polarization state and right-handed circular polarization state are used to encode the positive and negative values ​​of the weights. The positive value in the spot matrix corresponds to the left-handed circular polarization state, and the negative value in the spot matrix corresponds to the right-handed circular polarization state.

[0057] In some embodiments, the metasurface layer is used to perform a dot product operation on the incident image and the numerical values ​​represented in the spot matrix to obtain a characteristic image.

[0058] In some embodiments, the tunable metasurface structure may further include a first light-transmitting substrate layer and a second light-transmitting substrate layer, wherein the metasurface layer may be disposed on the first light-transmitting substrate layer; and the second light-transmitting substrate layer may be disposed on the second electrode layer.

[0059] Optionally, the first light-transmitting substrate layer and the second light-transmitting substrate layer are both made of glass.

[0060] See also Figure 2 , Figure 2 This is the second structural schematic diagram of the tunable metasurface structure provided in an embodiment of the present invention.

[0061] In this embodiment, the tunable metasurface structure includes a first light-transmitting substrate layer 241, a metasurface layer 210, a first electrode layer 221, a liquid crystal layer 230, a second electrode layer 222 and a second light-transmitting substrate layer 242 stacked in sequence.

[0062] Specifically, the supersurface layer 210 is arranged on the first light-transmitting substrate layer 241, the first electrode layer 221 is arranged on the supersurface layer 210, the liquid crystal layer 230 is arranged on the first electrode layer 221, the second electrode layer 222 is arranged on the liquid crystal layer 230, and the second light-transmitting substrate layer 242 is arranged on the second electrode layer 222.

[0063] In this embodiment, the polarization state of the incident light is left-handed circular polarization (LCP), and after being tuned together by the liquid crystal layer and the metasurface layer, light with left-handed circular polarization (LCP) and right-handed circular polarization (RCP) polarization states can be obtained.

[0064] This embodiment implements optical convolution through a metasurface, offloading the computationally expensive convolution operation to a high-speed, low-power optical device. The anisotropy of the liquid crystal at the optoelectronic front end can change the polarization of light.

[0065] The present invention also provides an imager based on an optical neural network, see Figure 3-Figure 4 , Figure 3 is a schematic diagram of the structure of an imager based on an optical neural network provided in an embodiment of the present invention, Figure 4 3 is a schematic diagram of the structure of the metasurface layer provided by an embodiment of the present invention. An imager based on an optical neural network may include a photodetector 320 and the above-mentioned tunable metasurface structure 310.

[0066] The tunable metasurface structure 310 is used to obtain a characteristic image after polarization multiplexing and spatial multiplexing of the light of the received image, and the photodetector 320 is used to receive the light of the characteristic image of the tunable metasurface structure 310 and form an image. The tunable metasurface structure 310 includes a liquid crystal module 311 and a metasurface module 312. The liquid crystal module 311 further includes the above-mentioned electrode layer and liquid crystal layer, and the metasurface module 312 includes the above-mentioned metasurface layer.

[0067] By controlling the thickness of the liquid crystal layer and controlling the orientation of the liquid crystal molecules in the liquid crystal layer by applying different voltages, a variable wave plate (VWP) with different phase delays can be realized. The VWP can continuously adjust the polarization state of the incident light, thereby changing the intensity of the two spiral channels incident on the metasurface module 312. Since the metasurface module 312 can realize polarization multiplexing of left-handed circular polarization and right-handed circular polarization, on the basis of the originally designed channel, the liquid crystal molecules realize nonlinear brightness changes in the PSF when the rotation angles are 0°, 20°, 35° and 45°, and the two circularly polarized lights in the optical convolution correspond to positive and negative values ​​in the digital convolution kernel respectively.

[0068] Nanocolumns are arranged in the supersurface layer of the tunable supersurface structure, and the nanocolumns serve as basic units in the supersurface layer.

[0069] In this embodiment, the nanopillars of different widths and lengths in the metasurface layer can also be optimized to act as phase-corresponding half-wave plates. By combining geometric phase modulation with local resonant phase delay, spin decoupling can be achieved, thereby realizing independent phase manipulation of orthogonal circularly polarized light.

[0070] In deep learning, convolution usually involves multiplying the target image with an A×A matrix, where each pixel in the matrix has an independent weight. This matrix is ​​applied to the image through a dot product operation and moved step by step to produce a single feature map.

[0071] In this embodiment, under the condition of incoherent illumination, PSY (x, y) is set to produce A × A foci, each with a specific intensity corresponding to the desired digital core weight. By adjusting the rotation of the liquid crystal molecules, the polarization state of the incoherent light can be modified to achieve nonlinear brightness adjustment of the foci. These foci produce overlapping A × A images on a sensor such as a photodetector, which are spatially moved according to their respective positions. Therefore, these weighted images are rasterized and summed by superimposing them on a sensor such as a photodetector.

[0072] The tunable metasurface structure achieves multi-channel signal processing through voltage regulation, rather than the convolution operation commonly performed in digital neural networks. The tunable metasurface structure consists of a voltage-regulated encapsulated liquid crystal layer and a pre-designed nuclear-encoded two-layer metasurface. The object is illuminated using a non-correlated light source, and the direction of the rotating liquid crystal is adjusted by voltage changes to achieve multi-dimensional optical convolution of the double helix channel, where LCP represents positive values ​​and RCP represents negative values. Polarization-sensitive photodetectors can capture polarization-related feature maps. Each photodetector pixel is equipped with a directional grating to achieve classification of polarization signals, thereby ensuring accurate feature extraction and efficient data processing.

[0073] See also Figure 5 , Figure 5 It is a schematic diagram of the principle of an imager based on an optical neural network provided in an embodiment of the present invention.

[0074] Figure 5 The comparison process of optical convolution and digital convolution is given. The input object is the light of the target to be identified with different intensities (Intensiyt). If it is processed by optical convolution (Optical), a spot matrix is ​​obtained, and finally the optical convolution result (Optical output) is output; if it is processed by digital convolution (Digital), a digital matrix is ​​obtained, and finally the digital convolution result (Digital output) is output. In optical convolution, each point in the spread function (PSF) has a brightness distribution corresponding to the weight of the digital convolution kernel.

[0075] The present invention also provides a target recognition method based on an optoelectronic hybrid convolutional neural network, see Figure 6-7 , Figure 6 is a flow chart of a target recognition method based on an optoelectronic hybrid convolutional neural network provided by an embodiment of the present invention, Figure 7 It is a structural schematic diagram of the optoelectronic hybrid convolutional neural network architecture provided by an embodiment of the present invention.

[0076] In this embodiment, the target recognition method based on the optoelectronic hybrid convolutional neural network uses the above-mentioned tunable metasurface structure. The target recognition method based on the optoelectronic hybrid convolutional neural network may include steps S610 to S630, and each step is specifically as follows:

[0077] S610: Perform convolution processing of an optical neural network on the target to be identified through a tunable metasurface structure to obtain a feature image after feature enhancement.

[0078] S620: Convert the characteristic image from an optical signal to an electrical signal through a photoelectric detector.

[0079] S630: Inputting the characteristic image of the electrical signal into the digital neural network for target recognition, and obtaining a recognition result output by the digital neural network.

[0080] Combination Figure 5 and Figure 7 This embodiment includes a tunable metasurface structure of a liquid crystal layer (LC) and a first metasurface layer (MS1) as an optoelectronic front end, which is a composite optical device (Compound meta-optic) including meta-optical functions, which can replace the convolution layer of a digital neural network. A photodetector (photodetector) including gratings (Gratings) as a digital back end can further identify feature images.

[0081] In this embodiment, the designed tunable metasurface structure serves as the optoelectronic front end, while the digital back end includes a digital maximum pooling module, a rectified linear unit activation function, and a fully connected layer. The optical convolution layer can use 12 designed convolution channels to generate feature maps, while the fixed convolution channel changes nonlinearly by adjusting the voltage. Each kernel consists of N=3 pixels, and the tunable metasurface structure introduces an additional voltage dimension to enhance object recognition in different scenarios.

[0082] For classification tasks, the feature map generated by the tunable metasurface structure is input into the back-end of the digital neural network for recognition, and the maximum probability can be selected from the preliminary recognition results as the recognition result.

[0083] In order to further illustrate the object recognition results of the embodiment of the present invention, some data validation sets are given below for illustration.

[0084] ① The dataset contains 60,000 training images, each with a resolution of 28×28 pixels. The performance of the system is evaluated by extracting feature maps of 10,000 MNIST images using an optical imager. The accuracy of digital convolution and optical convolution is 98.54% and 98.5%, respectively. The slight decrease in the accuracy of optical convolution may be due to the inaccuracy of phase optimization.

[0085] ②In order to further verify the robustness of the tunable metasurface structure, the Fashion MNIST dataset containing 60,000 training images of different clothing was used, whose spatial frequency was higher than that of the MNIST dataset in ①. The Fashion MNIST dataset was classified using a digital neural network and an optoelectronic neural network, with a digital classification accuracy of 91.0% and an optical classification accuracy of 90.9%. Compared with the MNIST dataset in ①, the accuracy of the recognition model processing the Fashion MNIST dataset is slightly lower, which is due to the presence of more complex feature information in the Fashion MNIST dataset. The training accuracy of the MNIST digital classification reference model without a convolutional layer is 80.1%, which highlights the key role of the convolution operation.

[0086] In summary, the embodiments of the present invention provide a tunable metasurface structure, an imager, and a target recognition method, wherein the tunable metasurface structure, as an optoelectronic front end, replaces the traditional optical imaging element, and can process various coded information, thereby improving the classification accuracy and robustness of various objects. The system realizes flexible tuning of two spiral channels by continuously switching voltages, and facilitates nonlinear brightness adjustment of multiple fixed information channels to generate multiple optical convolution channels; the positive and negative nuclei of left-handed circular polarization (LCP) and right-handed circular polarization (RCP) are incoherently irradiated by using the metasurface polarization multiplexing principle, and the controllability of the optoelectronic front end is realized; by integrating the tunable metasurface structure with a digital back end based on a photodetector / chip, etc., the speed can be increased, and data reading and transmission can be realized without an analog-to-digital converter, which will help achieve ultra-fast and low-latency effects. In addition, the optical convolution layer overcomes the limitations of the optical front end of the traditional imager in terms of depth or number of layers, paving the way for large-scale models in neural network applications. Therefore, this tunable metasurface structure not only generates rich information channels within a limited metasurface area but also provides flexible parameter adjustment, which has great potential in accelerating matrix operations.

[0087] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0088] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A tunable metasurface structure, characterized in that: It comprises a first electrode layer, a second electrode layer, a liquid crystal layer and a super surface layer; the first electrode layer is arranged on the super surface layer; the liquid crystal layer is arranged on the first electrode layer; the second electrode layer is arranged on the liquid crystal layer; Wherein, by applying voltage to the first electrode layer and the second electrode layer, the direction of the liquid crystal molecules in the liquid crystal layer is controlled to adjust the polarization state of the light in the incident image; the metasurface layer responds differently to light of different polarization states through polarization multiplexing, and the left-handed circular polarization state and the right-handed circular polarization state are used to encode the positive and negative values ​​of the weight, and a multi-channel A×A spot matrix is ​​generated through spatial multiplexing to determine the emitted characteristic image through the spot matrix; wherein A represents the number of rows and columns of the spot matrix; different spot brightness in the spot matrix is ​​used to characterize different convolution kernel weights; The tunable metasurface structure is expressed as: ； in, represents the operating wavelength; x and y represent the position information of the abscissa and ordinate in the plane of the metasurface layer respectively; represents the phase of the i-th spot in the spot matrix, represents the focal length of the metasurface layer, It represents the offset distance of the i-th spot in the spot matrix relative to the central spot in the x direction. Indicates the offset distance of the i-th spot in the spot matrix relative to the central spot in the y direction.

2. The tunable metasurface structure according to claim 1, characterized in that: The adjusting the polarization state of the light in the incident image comprises adjusting the ratio of the left-handed circular polarization state and the right-handed circular polarization state of the light in the incident image; In the optical neural network of the metasurface layer, the positive values ​​in the light spot matrix correspond to left-handed circular polarization states, and the negative values ​​in the light spot matrix correspond to right-handed circular polarization states.

3. The tunable metasurface structure according to claim 2, characterized in that: The properties of the liquid crystal layer are expressed as: ； in, Indicates the direction of rotation of liquid crystal molecules; represents the constant axis refractive index; represents the extraordinary refractive index; Represents the effective refractive index.

4. The tunable metasurface structure according to claim 2, characterized in that: The circularly polarized light multiplexing of the metasurface layer is expressed as: ； in, represents the rotation angle of the nanocolumns in the super surface layer; Indicates when When set to 0, the phase delay experienced by the nanopillars along the x direction in the metasurface layer; represents the phase delay of the left-handed circular polarization state; represents the phase delay of the right-hand circular polarization state.

5. The tunable metasurface structure according to claim 1, characterized in that: Also includes a first light-transmitting substrate layer and a second light-transmitting substrate layer; The super surface layer is arranged on the first light-transmitting substrate layer; and the second light-transmitting substrate layer is arranged on the second electrode layer.

6. The tunable metasurface structure according to claim 1, characterized in that: The metasurface layer is used to perform a dot product operation on the incident image and the numerical value represented in the light spot matrix to obtain the characteristic image.

7. The tunable metasurface structure according to claim 4, characterized in that: The thickness of the liquid crystal layer ranges from 1 micron to 10 microns, the cross-sectional shape of the nanocolumn is rectangular or elliptical; the length of the cross-sectional shape of the nanocolumn ranges from 100 nanometers to 1000 nanometers; the height of the nanocolumn ranges from 100 nanometers to 1000 nanometers; the distribution period of the nanocolumn ranges from 200 nanometers to 1000 nanometers.

8. The tunable metasurface structure according to claim 5, characterized in that: The materials of the first electrode layer and the second electrode layer are both indium tin oxide; The first light-transmitting substrate layer and the second light-transmitting substrate layer are both made of glass.

9. An imager based on an optical neural network, characterized in that: A photodetector and a tunable metasurface structure according to any one of claims 1 to 8; The tunable metasurface structure is used to perform polarization multiplexing and spatial multiplexing on the light of the received image to obtain a characteristic image, and the photodetector is used to receive the light of the characteristic image of the tunable metasurface structure and form an image.

10. A target recognition method based on an optoelectronic hybrid convolutional neural network, characterized in that: Using the tunable metasurface structure according to any one of claims 1 to 8, the target recognition method based on the optoelectronic hybrid convolutional neural network comprises: The tunable metasurface structure is used to perform convolution processing of an optical neural network on the target to be identified, so as to obtain a feature image after feature enhancement; Converting the characteristic image from an optical signal to an electrical signal through a photodetector; The characteristic image of the electrical signal is input into a digital neural network for target recognition, and a recognition result output by the digital neural network is obtained.

Images