Image classification method and image classification device
By using all-dielectric metasurface structures in image classification instead of hidden layers in traditional models, the problem of low accuracy and efficiency of traditional models in image classification is solved, higher image classification accuracy and efficiency are achieved, and system complexity and manufacturing costs are reduced.
Patent Information
- Application Number
- CN202411531609.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-30
- Publication Date
- 2025-02-28
AI Technical Summary
The traditional diffraction deep neural network model has problems such as low accuracy and efficiency, high system complexity and high manufacturing cost in image classification.
By training the diffraction deep neural network model, the phase modulation parameters of each hidden layer are extracted, the attribute parameters of the all-dielectric metasurface structure are determined, and the all-dielectric metasurface structure is used instead of the hidden layer in the traditional model to achieve precise control of the optical wave phase.
It improves the accuracy and efficiency of image classification, reduces system complexity and manufacturing costs, and the design of the all-dielectric metasurface structure is relatively simple and easy to make.
Smart Images
Figure CN119540611A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field related to image processing, and in particular to an image classification method and an image classification device. Background Art
[0002] At present, Diffractive Deep Neural Network (Diffractive Deep Neural Network, ) technology has shown its potential in processing large-scale image datasets in the field of image recognition, especially in tasks such as handwritten digit recognition and object classification. However, traditional Models often rely on specific optical component designs. Typically, existing The model uses different thicknesses to achieve phase control of light, and then uses 3D printing to make the diffraction layer. This design method not only increases the complexity of the system, but also reduces the accuracy and efficiency of image classification, and also has higher requirements for production precision, which increases the manufacturing cost. Summary of the invention
[0003] The present application aims to solve at least one of the technical problems existing in the prior art. To this end, the present application proposes an image classification method that can improve the accuracy and efficiency of image classification while reducing system complexity and manufacturing cost.
[0004] The present application also provides an image classification device, a control device for executing the above-mentioned image classification method, and a computer-readable storage medium.
[0005] According to the first aspect of the present application, the image classification method includes: Get the input image training set; A diffraction deep neural network model is obtained by training based on the input image training set; Extracting phase modulation parameters of each neuron in each hidden layer of the diffraction deep neural network model; Determine the attribute parameters of each unit structure of each sub-all-dielectric metasurface in the all-dielectric metasurface structure according to the phase modulation parameters of each neuron of each hidden layer in the diffraction deep neural network model, the number of layers of the sub-all-dielectric metasurface is the same as the number of layers of the hidden layer, and the neurons correspond one-to-one to the unit structures; The all-dielectric metasurface structure is manufactured according to the property parameters of each unit structure of each layer of the sub-all-dielectric metasurface; The all-dielectric metasurface structure is used for image recognition and image classification.
[0006] The image classification method according to the embodiment of the present application has at least the following beneficial effects: First, the diffraction deep neural network model is trained so that the model can fully learn the image features and perform correct classification. Then, the phase modulation parameters of each neuron in each hidden layer of the diffraction deep neural network model are extracted to determine the property parameters of each unit structure of each sub-all-dielectric metasurface in the all-dielectric metasurface structure. The high resolution, low loss and high transmittance characteristics of the all-dielectric metasurface structure are used to replace the hidden layer that uses thickness to modulate the phase in the traditional diffraction deep neural network model, so as to achieve precise control of the phase of the light wave, thereby improving the accuracy and efficiency of image classification. At the same time, since the structural design of the all-dielectric metasurface structure is relatively simple and easy to manufacture, the system complexity and manufacturing cost are reduced.
[0007] According to some embodiments of the present application, each of the sub-all-dielectric metasurfaces includes: a substrate structure, including a plurality of unit substrates; a plurality of nano cylinders corresponding to the plurality of unit substrates one by one, each of the nano cylinders is arranged on the upper surface of the corresponding unit substrate, each of the nano cylinders and the corresponding unit substrate constitute a unit structure, and the plurality of nano cylinders are arranged periodically; The property parameters include the radius of the nanocylinder; The method of determining the attribute parameters of each unit structure of each sub-all-dielectric metasurface in the all-dielectric metasurface structure according to the phase modulation parameters of each neuron in each hidden layer in the diffraction deep neural network model includes: Acquiring a corresponding relationship between the preset phase modulation parameter and the radius of the nanocylinder; According to the corresponding relationship and the phase modulation parameters of each of the neurons in each hidden layer in the diffraction deep neural network model, the radius of each of the nano-cylinders in each layer of the sub-all-dielectric supersurface in the all-dielectric supersurface structure is determined.
[0008] According to some embodiments of the present application, the number of layers of the sub-all-dielectric metasurface and the number of layers of the hidden layer are both three, and the number of neurons and the number of unit structures are both 32*32.
[0009] According to some embodiments of the present application, the size of the unit substrate is 360nm*360nm*350nm, the height of the nanocylinder is 180nm, the maximum radius of the nanocylinder is 80nm, and the minimum radius of the nanocylinder is 30nm.
[0010] According to some embodiments of the present application, the method of performing image recognition and image classification using the all-dielectric metasurface structure includes: Get the target input image; The target input image is subjected to image recognition and classification through the all-dielectric metasurface structure to obtain a target classification result.
[0011] According to some embodiments of the present application, the wavelength of the incident plane wave used for image recognition and classification through the all-dielectric metasurface structure is 600 nm.
[0012] According to the second aspect of the present application, the image classification device includes an all-dielectric metasurface structure obtained by the image classification method as described in the first aspect of the present application.
[0013] The image classification device according to the embodiment of the present application has at least the following beneficial effects: The image classification method of the first aspect of the embodiment is adopted. The diffraction deep neural network model is first trained so that the model can fully learn the image features and perform correct classification. Then, the phase modulation parameters of each neuron in each hidden layer in the diffraction deep neural network model are extracted, and the attribute parameters of each unit structure of each sub-all-dielectric metasurface in the all-dielectric metasurface structure are correspondingly determined. The high resolution, low loss and high transmittance characteristics of the all-dielectric metasurface structure are used to replace the hidden layer with thickness modulated phase in the traditional diffraction deep neural network model, so as to achieve precise control of the phase of the light wave, thereby improving the accuracy and efficiency of image classification. At the same time, since the structural design of the all-dielectric metasurface structure is relatively simple and easy to manufacture, the system complexity and manufacturing cost are reduced.
[0014] According to some embodiments of the present application, each of the sub-all-dielectric metasurfaces comprises: A substrate structure, comprising a plurality of unit substrates; A plurality of nano cylinders corresponding to the plurality of unit substrates one by one, each of the nano cylinders is arranged on the upper surface of the corresponding unit substrate, each of the nano cylinders and the corresponding unit substrate form a unit structure, and the plurality of nano cylinders are arranged periodically.
[0015] According to some embodiments of the present application, the material of the substrate structure is silicon dioxide.
[0016] According to some embodiments of the present application, the material of the nano cylinder is silicon.
[0017] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent from the description, or may be understood by practicing the present application. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which: Figure 1 is a flow chart of an image classification method according to an embodiment of the present application; Figure 2 is a schematic diagram of a diffraction deep neural network model according to an embodiment of the present application; Figure 3 is a three-dimensional diagram of an all-dielectric metasurface structure according to an embodiment of the present application; Figure 4 is a side view of an all-dielectric metasurface structure according to an embodiment of the present application; Figure 5 is a three-dimensional diagram of a unit structure of an embodiment of the present application; Figure 6 is a schematic diagram of the corresponding relationship between the phase modulation parameters and the radius of the nanocylinder according to an embodiment of the present application; Figure 7 This is a confusion matrix diagram of the accuracy of the numbers "0" to "5" of the trained diffraction deep neural network model in one embodiment of the present application; Figure 8 is a graph showing changes in accuracy and loss value of a diffraction deep neural network model in one embodiment of the present application during five hundred training batches; Fig. 9 This is a diagram showing the recognition result of the number “0” using an all-dielectric metasurface structure according to an embodiment of the present application; Fig.10 yes Fig. 9 Energy proportion diagram in each area in the recognition result; Fig.11 This is a diagram showing the recognition result of the number "1" using an all-dielectric metasurface structure according to an embodiment of the present application; Fig.12 yes Fig.11 Energy proportion diagram in each area in the recognition result; Fig.13 This is a diagram showing the recognition result of the number "2" using an all-dielectric metasurface structure according to an embodiment of the present application; Fig.14 yes Fig.13 Energy proportion diagram in each area in the recognition result; Fig.15 This is a diagram showing the recognition result of the number "3" using an all-dielectric metasurface structure according to an embodiment of the present application; Fig.16 yes Fig.15 Energy proportion diagram in each area in the recognition results.
[0019] Reference numerals: Incident plane wave 100; Input layer 200; Hidden layer 300; Output layer 400; Unit substrate 500; Nano cylinder 600. DETAILED DESCRIPTION
[0020] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as limiting the present application.
[0021] In the description of this application, if there is a description of first, second, etc., it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.
[0022] In the description of the present application, it should be understood that the descriptions involving orientation, such as the orientation or positional relationship indicated as up, down, etc., are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.
[0023] In the description of this application, it should be noted that, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meaning of the above terms in this application based on the specific content of the technical solution.
[0024] The following will be combined Figures 1 to 16 The image classification method according to the embodiment of the present invention is described clearly and completely. Obviously, the embodiment described below is only a part of the embodiments of the present invention, not all of the embodiments.
[0025] refer to Figure 1 , Figure 1 is a flowchart of an image classification method according to an embodiment of the present invention.
[0026] The image classification method according to the first aspect of the present invention includes: Get the input image training set; A diffraction deep neural network model is obtained by training based on the input image training set; Extracting phase modulation parameters of each neuron in each hidden layer 300 in the diffraction deep neural network model; Determine the attribute parameters of each unit structure of each sub-all-dielectric metasurface in the all-dielectric metasurface structure according to the phase modulation parameters of each neuron in each hidden layer 300 in the diffraction deep neural network model, the number of layers of the sub-all-dielectric metasurface is the same as the number of layers of the hidden layer 300, and the neurons correspond to the unit structures one by one; The all-dielectric metasurface structure is manufactured according to the property parameters of each unit structure of each layer of the sub-all-dielectric metasurface; Image recognition and image classification using all-dielectric metasurface structures.
[0027] refer to Figure 2 , Figure 2 Schematic diagram of a diffraction deep neural network model of an embodiment of the present application. The diffraction deep neural network model includes an incident plane wave 100, an input layer 200, a hidden layer 300 and an output layer 400. During the training process of the diffraction deep neural network model, data enters the network through the input layer 200 and propagates forward to generate output, and the weights of the neurons are optimized by the error back propagation algorithm. According to the Rayleigh-Sommerfeld diffraction formula, each point on the diffraction surface can be regarded as a secondary spherical wavelet source, and the spatial distribution of the light field is the result of the coherent superposition of the wavelet sources after diffraction propagation. The diffraction propagation formula of the neuron is expressed as: ;Formula (1) in, l represents the diffractive neural network l layer, i Indicates l Layer i neurons, r Indicates l Layer i The neurons and l +1) Euclidean distance between layers, j Representing imaginary units , represents the wavelength of the incident plane wave 100, Indicates the first l +1) layer is located at l The coordinates in the layer are No. i The complex-valued field generated by each neuron, the input layer 200 is designated as layer 0. As a secondary wave source, the amplitude and relative phase of the secondary waves generated by the neurons are determined by the incident plane wave 100 and the transfer function of the neurons. l Tier i The output light field of a neuron can be expressed as: ;Formula (2) in, represents the amplitude of the neuron, represents the phase modulation parameter of the neuron.
[0028] In actual manufacturing, the selected substrate structure is usually optically transparent and has a constant transmittance. The neurons processed with this substrate structure have a constant amplitude (amplitude is used to represent transmittance). Therefore, it is feasible to realize image classification with an all-dielectric metasurface structure. By using the above unit structure and determining the attribute parameters based on the phase modulation parameters obtained by the diffraction deep neural network model, and arranging them according to the attribute parameters, the phase of light can be accurately modulated, and finally the correct classification of the input image is achieved, and it has the characteristics of low loss and high transmittance.
[0029] In some embodiments, the input image training set includes multiple input images, which are pre-processed and then used as inputs into the diffraction deep neural network model to simulate the process of light passing through several hidden layers 300, and then the diffraction deep neural network model is trained using a neural network algorithm so that the diffraction deep neural network model can fully learn the image features and correctly classify (i.e., the energy is relatively maximum in the set area corresponding to the number). Figure 8 As shown, Figure 8 is a graph showing the change in accuracy and loss value of the diffraction deep neural network model in one embodiment of the present application during five hundred training batches. Figure 8 It can be seen that the diffraction deep neural network model converges quickly during the training process, indicating that the model has been effectively optimized and adjusted. Then, the phase modulation parameters of each neuron on each hidden layer 300 are extracted from the diffraction deep neural network model, and the property parameters of the unit structure at each position of the all-dielectric metasurface structure can be obtained according to the phase modulation parameters.
[0030] In some embodiments of the present application, reference Figures 3 to 6 , Figure 3 is a three-dimensional diagram of an all-dielectric metasurface structure according to an embodiment of the present application, Figure 4 is a side view of an all-dielectric metasurface structure according to an embodiment of the present application, Figure 5 is a three-dimensional diagram of a unit structure of an embodiment of the present application, Figure 6 is a schematic diagram of the corresponding relationship between the phase modulation parameters and the radius of the nano cylinder 600 according to an embodiment of the present application, Figure 6The horizontal axis in is the radius of the nano cylinder 600, the vertical axis on the left is the phase modulation parameter, and the vertical axis on the right is the amplitude. Each sub-all-dielectric metasurface includes: a substrate structure, including a plurality of unit substrates 500; a plurality of nano cylinders 600 corresponding to the plurality of unit substrates 500, each nano cylinder 600 is arranged on the upper surface of the corresponding unit substrate 500, each nano cylinder 600 and the corresponding unit substrate 500 form a unit structure, and the plurality of nano cylinders 600 are arranged periodically; the property parameters include the radius of the nano cylinder 600; The attribute parameters of each unit structure of each sub-all-dielectric metasurface in the all-dielectric metasurface structure are determined according to the phase modulation parameters of each neuron in each hidden layer 300 in the diffraction deep neural network model, including: Obtaining a corresponding relationship between a preset phase modulation parameter and a radius of the nano cylinder 600; According to the corresponding relationship and the phase modulation parameters of each neuron in each hidden layer 300 in the diffraction deep neural network model, the radius of each nanocylinder 600 of each sub-all-dielectric metasurface in the all-dielectric metasurface structure is determined.
[0031] The embodiment of the present application can adjust the radius of each nano cylinder 600 to achieve 0-2 The phase is precisely modulated, and has the characteristics of low loss and high transmittance. The all-dielectric metasurface structure is used to replace the traditional hidden layer 300 that modulates the phase by thickness, which can improve the accuracy and efficiency of image classification, and the structural design is relatively simple, which reduces the manufacturing cost.
[0032] It should be noted that the corresponding relationship between the phase modulation parameter and the radius of the nano cylinder 600 is obtained through experiments, and the specific experimental principle and process will not be described in detail here.
[0033] In addition, it should be noted that the radius of each nano-cylinder 600 of each layer of the sub-all-dielectric metasurface may be the same or different, which is determined according to the phase modulation parameters of the neurons actually calculated. The radius of the nano-cylinder 600 can also be obtained by comprehensive calculation based on the phase modulation parameters and amplitude, and the specific calculation method is not specifically limited here.
[0034] In some embodiments of the present application, the number of layers of the sub-all-dielectric metasurface and the number of layers of the hidden layer 300 are both three, and the number of neurons and the number of unit structures are both 32*32. The three-layer 32*32 structure is adopted in the embodiment of the present application in order to take into account both small volume and high recognition accuracy. It is only a special case, and other layout methods can be selected according to actual needs, which cannot be regarded as a limitation of the present application.
[0035] In some embodiments of the present application, the size of the unit substrate 500 is 360nm*360nm*350nm, the height of the nano cylinder 600 is 180nm, the maximum radius of the nano cylinder 600 is 80nm, and the minimum radius of the nano cylinder 600 is 30nm.
[0036] In some embodiments of the present application, the wavelength of the incident plane wave 100 used for image recognition and classification by the all-dielectric metasurface structure is 600 nm.
[0037] It should be noted that the above-mentioned size of the unit substrate 500, the height of the nanocylinder 600, the maximum and minimum values of the radius of the nanocylinder 600, and the wavelength of the incident plane wave 100 are all special cases, and other specific values can be selected according to actual needs, and cannot be regarded as limitations on the present application.
[0038] In some embodiments of the present application, reference Figure 7 as well as Figures 9 to 16 , Figure 7 This is a confusion matrix diagram of the accuracy of the numbers "0" to "5" of the trained diffraction deep neural network model in one embodiment of the present application. Fig. 9 This is a diagram showing the recognition result of the number “0” using an all-dielectric metasurface structure according to an embodiment of the present application. Fig.10 yes Fig. 9 The energy proportion diagram in each area in the recognition result, Fig.11 This is a diagram showing the recognition result of the number "1" using an all-dielectric metasurface structure in one embodiment of the present application. Fig.12 yes Fig.11 The energy proportion diagram in each area in the recognition result, Fig.13 This is a diagram showing the recognition result of the number "2" using an all-dielectric metasurface structure according to an embodiment of the present application. Fig.14 yes Fig.13 The energy proportion diagram in each area in the recognition result, Fig.15 This is a diagram showing the recognition result of the number "3" using an all-dielectric metasurface structure in one embodiment of the present application. Fig.16 yes Fig.15 Energy proportion diagram in each area in the recognition results.
[0039] Image recognition and image classification using all-dielectric metasurface structures, including: Get the target input image; The target input image is subjected to image recognition and classification through the all-dielectric metasurface structure to obtain the target classification result.
[0040] Figure 7The relationship between the actual category and the predicted category for image recognition and classification through the all-dielectric metasurface structure is shown. Through this matrix, we can understand the accuracy of recognizing the numbers "0" to "5" through the all-dielectric metasurface structure, as well as its specific performance on each number, including which numbers are easily misclassified and which numbers they are usually misidentified as. Figure 7 In the figure, the horizontal axis represents the predicted category and the vertical axis represents the actual category. For example, the horizontal axis is "2" and the vertical axis is "1", and the corresponding number "0.007" represents the error rate of identifying the number "1" as "2"; the number "0.954" corresponding to both the horizontal and vertical axes being "2" represents the accuracy rate of identifying the number "2" through the all-dielectric metasurface structure.
[0041] In some embodiments, from Figures 9 to 16 It can be seen that the numbers "0" to "3" can be correctly identified using the all-dielectric metasurface structure.
[0042] According to the image classification method of the embodiment of the present application, the diffraction deep neural network model is first trained to enable the model to fully learn the image features and perform correct classification, and then the phase modulation parameters of each neuron in each hidden layer 300 in the diffraction deep neural network model are extracted, and the attribute parameters of each unit structure of each sub-all-dielectric metasurface in the all-dielectric metasurface structure are correspondingly determined. The high resolution, low loss and high transmittance characteristics of the all-dielectric metasurface structure are used to replace the hidden layer 300 with thickness modulated phase in the traditional diffraction deep neural network model, so as to achieve precise control of the phase of the light wave, thereby improving the accuracy and efficiency of image classification. At the same time, since the structural design of the all-dielectric metasurface structure is relatively simple and easy to manufacture, the system complexity and manufacturing cost are reduced.
[0043] According to the second aspect of the present application, the image classification device includes an all-dielectric metasurface structure obtained by the image classification method of the first aspect of the present application.
[0044] According to the image classification device of the embodiment of the present application, the image classification method of the first aspect of the embodiment is adopted. The diffraction deep neural network model is first trained so that the model can fully learn the image features and perform correct classification. Then, the phase modulation parameters of each neuron in each hidden layer 300 in the diffraction deep neural network model are extracted, and the attribute parameters of each unit structure of each sub-all-dielectric metasurface in the all-dielectric metasurface structure are correspondingly determined. The high resolution, low loss and high transmittance characteristics of the all-dielectric metasurface structure are used to replace the hidden layer 300 with thickness modulated phase in the traditional diffraction deep neural network model, so as to achieve precise control of the phase of the light wave, thereby improving the accuracy and efficiency of image classification. At the same time, since the structural design of the all-dielectric metasurface structure is relatively simple and easy to manufacture, the system complexity and manufacturing cost are reduced.
[0045] In some embodiments of the present application, reference Figures 3 to 5 , each sub-all-dielectric metasurface includes: A substrate structure, comprising a plurality of unit substrates 500; A plurality of nano cylinders 600 correspond one to one with a plurality of unit substrates 500 , each nano cylinder 600 is disposed on the upper surface of the corresponding unit substrate 500 , each nano cylinder 600 and the corresponding unit substrate 500 form a unit structure, and the plurality of nano cylinders 600 are arranged periodically.
[0046] The embodiment of the present application can adjust the radius of each nano cylinder 600 to achieve 0-2 The phase is precisely modulated, and has the characteristics of low loss and high transmittance. The all-dielectric metasurface structure is used to replace the traditional hidden layer 300 that modulates the phase by thickness, which can improve the accuracy and efficiency of image classification, and the structural design is relatively simple, which reduces the manufacturing cost.
[0047] In some embodiments of the present application, the material of the base structure is silicon dioxide. The material of the nano cylinder 600 is silicon. Silicon dioxide and silicon have high transmittance, which can improve the signal-to-noise ratio of the photon computing system. In contrast, traditional thickness modulated diffraction elements usually require the use of multi-layer material stacking, resulting in reduced transmittance and low image recognition and classification accuracy.
[0048] It should be noted that the base structure and the nano-cylinders 600 may also be made of other high-transmittance materials, which should not be regarded as a limitation to the present application.
[0049] It will be appreciated by those skilled in the art that all or some of the steps and systems in the methods disclosed above may be implemented as software, firmware, hardware, and appropriate combinations thereof. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or transient medium). As known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.
[0050] The embodiments of the present application are described in detail above in conjunction with the accompanying drawings, but the present application is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in the relevant technical field without departing from the purpose of the present application.
Claims
1. An image classification method, characterized in that: include: Get the input image training set; A diffraction deep neural network model is obtained by training based on the input image training set; Extracting phase modulation parameters of each neuron in each hidden layer of the diffraction deep neural network model; Determine the attribute parameters of each unit structure of each sub-all-dielectric metasurface in the all-dielectric metasurface structure according to the phase modulation parameters of each neuron of each hidden layer in the diffraction deep neural network model, the number of layers of the sub-all-dielectric metasurface is the same as the number of layers of the hidden layer, and the neurons correspond one-to-one to the unit structures; The all-dielectric metasurface structure is manufactured according to the property parameters of each unit structure of each layer of the sub-all-dielectric metasurface; The all-dielectric metasurface structure is used for image recognition and image classification.
2. The image classification method according to claim 1, characterized in that: Each of the sub-all-dielectric metasurfaces comprises: a substrate structure, comprising a plurality of unit substrates; a plurality of nano cylinders corresponding to the plurality of unit substrates one by one, each of the nano cylinders being arranged on the upper surface of the corresponding unit substrate, each of the nano cylinders and the corresponding unit substrate forming a unit structure, and the plurality of nano cylinders being arranged periodically; The property parameters include the radius of the nanocylinder; The method of determining the attribute parameters of each unit structure of each sub-all-dielectric metasurface in the all-dielectric metasurface structure according to the phase modulation parameters of each neuron in each hidden layer in the diffraction deep neural network model includes: Acquiring a corresponding relationship between the preset phase modulation parameter and the radius of the nanocylinder; According to the corresponding relationship and the phase modulation parameters of each of the neurons in each hidden layer in the diffraction deep neural network model, the radius of each of the nano-cylinders in each layer of the sub-all-dielectric supersurface in the all-dielectric supersurface structure is determined.
3. The image classification method according to claim 1 or 2, characterized in that: The number of layers of the sub-all-dielectric metasurface and the number of layers of the hidden layer are both three, and the number of neurons and the number of unit structures are both 32*32.
4. The image classification method according to claim 2, characterized in that: The size of the unit substrate is 360nm*360nm*350nm, the height of the nano cylinder is 180nm, the maximum value of the radius of the nano cylinder is 80nm, and the minimum value of the radius of the nano cylinder is 30nm.
5. The image classification method according to claim 1, characterized in that: The method of using the all-dielectric metasurface structure to perform image recognition and image classification includes: Get the target input image; The target input image is subjected to image recognition and classification through the all-dielectric metasurface structure to obtain a target classification result.
6. The image classification method according to claim 5, characterized in that: The wavelength of the incident plane wave used for image recognition and classification through the all-dielectric metasurface structure is 600nm.
7. An image classification device, characterized in that: It includes an all-dielectric metasurface structure obtained by the image classification method as described in any one of claims 1 to 6.
8. The image classification device according to claim 7, characterized in that: Each of the sub-all-dielectric metasurfaces comprises: A substrate structure, comprising a plurality of unit substrates; A plurality of nano cylinders corresponding to the plurality of unit substrates one by one, each of the nano cylinders is arranged on the upper surface of the corresponding unit substrate, each of the nano cylinders and the corresponding unit substrate form a unit structure, and the plurality of nano cylinders are arranged periodically.
9. The image classification device according to claim 8, characterized in that: The material of the base structure is silicon dioxide.
10. The image classification device according to claim 8, characterized in that: The material of the nano cylinder is silicon.