Construction method of photoelectric computing chip based on metasurface and photoelectric computing chip

Abstract

The invention provides a construction method of a photoelectric computing chip based on a metasurface and the photoelectric computing chip, and relates to the technical field of photoelectric computing architecture. The invention discloses a construction method of a photoelectric computing chip based on a metasurface. The construction method comprises the following steps: acquiring a preset training data set; constructing a network model of the photoelectric computing chip, wherein the network model comprises a metasurface optical network and an electrical network; training and testing the network model through a preset training data set, and obtaining a test result; and under the condition that the test result meets a preset condition, constructing the photoelectric calculation chip through the network model. According to the embodiment of the invention, the image information can be quickly processed, and the energy consumption and reasoning time of a chip are reduced.

Description

Technical Field

[0001] The present application relates to the technical field of optoelectronic computing architecture, and more specifically, to a method for constructing an optoelectronic computing chip based on a metasurface and an optoelectronic computing chip. Background Art

[0002] With the advent of the big data era, intelligent recognition technology based on artificial intelligence has been applied in many fields, such as autonomous driving, medical care, finance, education, security, etc. It can replace humans to process large amounts of image information and perform target recognition.

[0003] At present, most of the algorithm models suitable for intelligent recognition data processing are still based on the traditional image recognition architecture (i.e. sensor-register-processor). When processing large scenes and multi-target image tasks, this structure has high requirements for transmission bandwidth, rate, computing power, etc., and most of the redundant data will still be transmitted and calculated, which greatly reduces the recognition efficiency and accuracy.

[0004] In order to solve the problems of high training energy consumption, slow inference speed and low recognition accuracy in the process of processing large image data of large scenes and multiple targets, existing solutions include image recognition systems based on large models and small algorithms based on yolo. Among them, the image recognition system based on large models needs to use huge training resources to train parameters of more than 100 million. Although it can achieve higher recognition accuracy, the huge model is difficult to deploy on a small terminal, and the inference speed is slow, and it is impossible to recognize the target quickly and in real time. Although the small algorithm based on yolo can recognize objects quickly and in real time, it has low efficiency and poor accuracy for large-scale image recognition. Therefore, it is difficult to achieve low-energy, high-efficiency, fast and real-time target recognition in large scenes and multi-target tasks based on the traditional image recognition architecture.

[0005] The inventors of this application discovered that diffraction light computing based on metasurfaces has the advantages of high-speed and high-parallel image processing. However, due to the non-reconfigurable physical structure, small processing scale, and difficulty in large-scale simulation, its task is single, the image recognition scenario is small, and the design of large-scale diffraction networks is difficult. It still cannot be used for large-scale intelligent recognition of general images. Summary of the invention

[0006] According to one aspect of the present application, a method for constructing an optoelectronic computing chip based on a metasurface is provided, comprising: obtaining a preset training data set; constructing a network model of the optoelectronic computing chip, the network model comprising a metasurface optical network and an electrical network; training and testing the network model using the preset training data set, and obtaining the test results; and constructing the optoelectronic computing chip using the network model when the test results meet preset conditions.

[0007] According to some embodiments, constructing a network model of an optoelectronic computing chip includes: performing metasurface simulation modeling of a metasurface optical network; setting parameters of the metasurface optical network; and setting parameters of an electrical network.

[0008] According to some embodiments, a metasurface simulation modeling of a metasurface optical network is performed, including: constructing a metasurface of the metasurface optical network through a preset structure, the metasurface of the metasurface optical network comprising a preset number of neurons; scanning the metasurface of the metasurface optical network through a preset wavelength to obtain a correspondence between a preset number of neurons and a phase.

[0009] According to some embodiments, a metasurface optical network includes at least one convolutional metasurface layer and at least one pooling metasurface layer; setting parameters of the metasurface optical network includes: configuring at least one convolutional metasurface layer and at least one pooling metasurface layer through a preset metasurface layout and a preset array scale; configuring at least one pooling metasurface layer through a preset distribution rule.

[0010] According to some embodiments, setting parameters of the electrical network includes: configuring an activation function of the electrical network according to a category corresponding to a preset training data set for image classification and recognition.

[0011] According to some embodiments, the network model is trained and tested using a preset training data set, and the test results are obtained, including: weight training of the metasurface optical network using a preset training data set to obtain the weight distribution of the metasurface optical network; accuracy testing of the metasurface optical network based on the weight distribution of the metasurface optical network, and obtaining a first test result.

[0012] According to some embodiments, training and testing the network model through a preset training data set and obtaining test results also includes: weight training the electrical network through a preset training data set to obtain the weight distribution of the electrical network; performing an accuracy test on the electrical network based on the weight distribution of the electrical network, and obtaining a second test result.

[0013] According to some embodiments, when the test results meet preset conditions, an optoelectronic computing chip is constructed through a network model, including: when the accuracy of the first test result and the second test result both meet the preset conditions, arraying the metasurface optical network to form at least one convolutional metasurface layer array and at least one pooling metasurface layer array; constructing an electrical network to form an optoelectronic computing chip with at least one convolutional metasurface layer array and at least one pooling metasurface layer array.

[0014] According to one aspect of the present application, an optoelectronic computing chip is provided, comprising: a general diffraction computing module, comprising at least one convolutional supersurface layer array and at least one pooling supersurface layer array, for performing feature extraction of image information contained in a received optical signal through at least one convolutional supersurface layer array, and for performing feature compression of the image information that has undergone feature extraction through at least one pooling supersurface layer array; an image sensing module, comprising an image sensor and an electrical network in a network model of an optoelectronic computing chip constructed by the aforementioned method, for converting an optical signal that has undergone feature extraction and feature compression of image information into an electrical signal through the image sensor, and for classifying and identifying the image information in the electrical signal through the electrical network; wherein, at least one convolutional supersurface layer array and at least one pooling supersurface layer array are formed by the aforementioned method, and at least one convolutional supersurface layer array and at least one pooling supersurface layer array correspond to each other.

[0015] According to one aspect of the present application, a method for intelligent image recognition is provided, comprising: acquiring an optical signal containing image information; processing the optical signal by means of an optoelectronic computing chip as described above, and acquiring a processing result; and determining a target to be recognized in the image information based on the processing result.

[0016] According to the embodiments of the present application, optoelectronic intelligent computing can be applied to realize intelligent image recognition of large scenes and multiple targets. By constructing a universal diffraction calculation module to process image information, the energy consumption and reasoning time of the chip are reduced. Through multi-layer metasurface unit design verification and flexible arraying, the difficulty of simulation design is reduced, the simulation design of large-scale diffraction metasurfaces is realized, and the reliability and versatility of the chip are improved.

[0017] It should be understood that the foregoing general description and the following detailed description are exemplary only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application.

[0019] Figure 1 A schematic diagram of the architecture of an optoelectronic computing chip according to an exemplary embodiment of the present application is shown.

[0020] Figure 2 A schematic diagram showing the calculation principle of a general diffraction calculation module according to an exemplary embodiment of the present application is shown.

[0021] Figure 3 A flow chart of a method for constructing an optoelectronic computing chip according to an exemplary embodiment of the present application is shown.

[0022] Figure 4 A schematic diagram showing the generation of a basic graphic data set according to an exemplary embodiment of the present application is shown.

[0023] Figure 5 A phase distribution diagram of a convolution metasurface layer according to an exemplary embodiment of the present application is shown.

[0024] Figure 6 A confusion matrix diagram showing the accuracy of a metasurface optical network recognizing a basic graphic data set according to an exemplary embodiment of the present application.

[0025] Figure 7 A confusion matrix diagram showing the accuracy of an electrical network recognizing a handwritten digital picture dataset according to an exemplary embodiment of the present application is shown.

[0026] Figure 8 A flowchart of an image intelligent recognition method according to an exemplary embodiment of the present application is shown. DETAILED DESCRIPTION

[0027] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this application will be comprehensive and complete and will fully convey the concepts of the example embodiments to those skilled in the art. The same reference numerals in the figures represent the same or similar parts, and thus their repeated description will be omitted.

[0028] The described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present application. However, those skilled in the art will appreciate that the technical scheme of the present application may be put into practice without one or more of these specific details, or other modes, components, materials, devices or operations may be adopted. In these cases, known structures, methods, devices, realizations, materials or operations will not be shown or described in detail.

[0029] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0030] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.

[0031] The present application provides a method for constructing an optoelectronic computing chip based on a metasurface and an optoelectronic computing chip, which can effectively and quickly perform real-time detection of large scene images and multi-target tasks, reduce reasoning energy consumption and rate, and improve multi-target recognition accuracy under high-speed reasoning.

[0032] A method for constructing an optoelectronic computing chip based on a metasurface and an optoelectronic computing chip according to an embodiment of the present application will be described in detail below with reference to the accompanying drawings.

[0033] Figure 1 A schematic diagram of the architecture of an optoelectronic computing chip according to an exemplary embodiment of the present application is shown.

[0034] like Figure 1 As shown, the architecture of the optoelectronic computing chip includes, from top to bottom, a convolutional supersurface layer array 100, a pooling supersurface layer array 200 and an image sensing module 300, wherein the convolutional supersurface layer array 100 and the pooling supersurface layer array 200 constitute a general diffraction computing module.

[0035] The convolutional metasurface layer array 100 includes at least one convolutional metasurface layer array, which is used to extract features from image information contained in an optical signal. Each convolutional metasurface layer array is composed of M×N convolutional metasurfaces. M and N are positive integers.

[0036] For example, the optoelectronic computing chip receives an optical signal, and the optical signal enters the general diffraction computing module and first passes through the convolution metasurface layer array 100. At least one convolution metasurface layer array in the convolution metasurface layer array 100 performs multiple convolution calculations on the image information contained in the optical signal to complete the feature extraction of the image.

[0037] The pooling super surface layer array 200 includes at least one pooling super surface layer array, which is used to compress the features extracted by the convolution super surface layer array 100. Each pooling super surface layer array is composed of M×N pooling super surfaces. M and N are positive integers.

[0038] For example, the optical signal passes through the convolution super surface layer array 100 and then enters the pooling super surface layer array 200. At least one pooling super surface layer array in the pooling super surface layer array 200 performs compression calculation on the extracted image features and obtains a small-scale feature compression matrix.

[0039] According to some embodiments, the general diffraction calculation module composed of the convolution metasurface layer array 100 and the pooling metasurface layer array 200 can adopt any one of semiconductor materials, conductor materials or insulator materials, such as gallium nitride (GaN), and be prepared by CMOS process.

[0040] The image sensing module 300 includes an image sensor for receiving an optical signal that has passed through a general diffraction calculation module and completed image feature extraction and compression, and converting the optical signal into an electrical signal through photoelectric conversion.

[0041] According to some embodiments, the image sensing module 300 further includes a constructed electrical network for processing the electrical signal after the photoelectric conversion to perform image classification and / or target recognition.

[0042] Understandably, Figure 1 The number of convolutional super surface layer arrays 100, pooling super surface layer arrays 200, and image sensing modules 300 is only for illustration. According to the actual scene requirements, there can be any number of convolutional super surface layer arrays 100, pooling super surface layer arrays 200, and image sensing modules 300.

[0043] Figure 2 A schematic diagram showing the calculation principle of a general diffraction calculation module according to an exemplary embodiment of the present application is shown.

[0044] like Figure 2 As shown, the general diffraction calculation module includes at least one convolution super surface layer array and at least one pooling super surface layer array.

[0045] According to some embodiments, the convolutional hypersurface layer array and the pooling hypersurface layer array correspond to each other, and the neurons in the convolutional hypersurface layer array and the neurons in the pooling hypersurface layer array correspond to each other.

[0046] The convolutional metasurface layer array is used to extract feature information in large scene images through convolution calculations.

[0047] The pooling metasurface layer array is used to compress and calculate the feature information of the extracted large scene image.

[0048] According to some embodiments, the convolutional metasurface layer array and the pooling metasurface layer array are formed by arraying the metasurface optical network of the optoelectronic computing chip that has undergone image feature extraction and image feature compression testing.

[0049] The general diffraction calculation module can divide large scene images into regions according to the training results of the metasurface optical network (such as Figure 2 Region 1, region 2 and region 3 as shown), and realize the extraction and compression of regional features of large scene images.

[0050] like Figure 2 As shown, the general diffraction calculation module extracts features of regions 1, 2, and 3 in the current image through the convolution metasurface layer array.

[0051] Furthermore, the general diffraction calculation module compresses the extracted features in region 1, region 2 and region 3 through the pooling metasurface layer array to form a feature compression matrix corresponding to the current image.

[0052] After the general diffraction calculation module generates the feature compression matrix information, the image sensing module in the optoelectronic computing chip receives the optical signal containing the feature compression matrix information and performs photoelectric conversion to classify and identify the obtained electrical signal through the electrical network and obtain the identification result.

[0053] Figure 3 A flow chart of a method for constructing an optoelectronic computing chip according to an exemplary embodiment of the present application is shown.

[0054] like Figure 3 As shown, in step S100, a preset training data set is obtained.

[0055] The preset training datasets include basic graphic datasets and handwritten digit image datasets.

[0056] Get basic graph datasets for training metasurface optical networks in network models of optoelectronic computing chips.

[0057] According to some embodiments, the basic graphic data set includes a primitive graphic data set, which includes graphics such as a straight line, a circle, a triangle, a rectangle, a hexagon, a cross, a circular ring, a right angle, a semicircular ring, and a rectangular ring.

[0058] According to some embodiments, the basic graphic data set also includes a data enhancement data set corresponding to the original graphic data set. The data enhancement performed on the original graphic data set includes scaling, shifting, rotating, and other processing of the images in the original graphic data set. Figure 4 As shown, to form a data augmentation dataset corresponding to the original graph dataset.

[0059] Obtain a dataset of handwritten digit images for training electrical networks in network models of optoelectronic computing chips.

[0060] In step S200, a network model of the optoelectronic computing chip is constructed.

[0061] According to some embodiments, the network model of the optoelectronic computing chip includes a metasurface optical network and an electrical network, wherein the metasurface optical network includes at least one convolutional metasurface layer and at least one pooling metasurface layer.

[0062] Perform metasurface simulation modeling of metasurface optical networks.

[0063] According to some embodiments, a metasurface of a metasurface optical network can be constructed with a preset structure through a preset simulation tool (e.g., Lumerical, COMSOL), so that the neurons included in the metasurface can change the phase and conform to the phase variation range. The phase variation range can be set to [0, 2π].

[0064] According to some embodiments, the metasurface may be of any shape. In an embodiment of the present application, the metasurface is a cylindrical structure based on gallium nitride (GaN) material, each cylindrical structure corresponds to a neuron, and each metasurface includes a preset number of neurons.

[0065] According to some embodiments, the metasurface may be scanned at a preset wavelength to obtain the corresponding relationship between a preset number of neurons and the phase. In an embodiment of the present application, a diameter scan may be performed at a preset wavelength to obtain the relationship between the diameter change of each cylindrical structure and the phase corresponding to the neuron. The preset wavelength may be set to 532 nm.

[0066] Furthermore, parameters of the metasurface optical network are set.

[0067] According to some embodiments, at least one convolutional metasurface layer and at least one pooling metasurface layer in a metasurface optical network can be configured by a preset metasurface layout and a preset array scale. In an embodiment of the present application, the metasurface optical network can be configured to have two convolutional metasurface layers and one pooling metasurface layer.

[0068] For example, each convolution hypersurface of the convolution hypersurface layer and each pooling hypersurface of the pooling hypersurface layer can be set according to a hypersurface layout scale of 1000×1000, wherein each convolution hypersurface and each pooling hypersurface contain 1000×1000 neurons.

[0069] For another example, the array of the convolutional hypersurface layer and the pooling hypersurface layer may be set to an array size of M×N, where the values ​​of M and N may be 100. Each convolutional hypersurface layer and each pooling hypersurface layer includes 100×100 hypersurfaces.

[0070] According to some embodiments, each pooling hypersurface in the pooling hypersurface layer may be set according to a preset distribution rule.

[0071] For example, it can be set according to the distribution rule of the off-axis focusing metasurface, and its distribution rule can be expressed by the following formula.

[0072] Φ(x,y)=(2π / λ)[f-((x f ) 2 +(yy f ) 2 +z f 2 )] 1 / 2 , (1)

[0073] Among them, Φ(x,y) is the phase value, x and y are the coordinate values, λ is the preset wavelength, and x f ,y f 、z f They are the offset of the focus in the x and y planes and the focal length of the z axis.

[0074] In an embodiment of the present application, a pooling super-surface layer set according to a preset distribution rule can perform off-axis compression on the extracted image features and obtain a 10×10 feature matrix.

[0075] Furthermore, the parameters of the electrical network are set.

[0076] According to some embodiments, the activation function of the electrical network can be configured according to the category corresponding to the preset handwritten digital picture data set, so that the electrical network can perform image classification and recognition.

[0077] For example, a handwritten digit image dataset includes 10 categories. The electrical network can be set to perform image classification and recognition of the 10 categories corresponding to the handwritten digit image dataset based on a softmax activation function.

[0078] In step S300, the network model is trained and tested using a preset training data set, and the test results are obtained.

[0079] Firstly, the weights of the metasurface optical network are trained through the basic graph dataset to obtain the weight distribution of the metasurface optical network.

[0080] According to some embodiments, weight training of a convolutional metasurface layer in a metasurface optical network may be performed via a back-propagating fully connected neural network.

[0081] In an embodiment of the present application, the metasurface optical network includes two convolutional metasurface layers, namely, a first convolutional metasurface layer and a second convolutional metasurface layer. Based on a basic graphic data set, the first convolutional metasurface layer and the second convolutional metasurface layer are trained by a fully connected neural network using back propagation to obtain the phase distribution corresponding to the first convolutional metasurface layer and the second convolutional metasurface layer, such as Figure 5As shown. The horizontal and vertical axes represent the index values ​​of the hypersurface rows and columns, which are consistent with the size of each hypersurface in the convolutional hypersurface layer. The legend indicates the phase size calculated at that position, in radians, ranging from 0 to 6.28.

[0082] According to some embodiments, the weight distribution of the metasurface optical network can be determined by convolving the phase distribution corresponding to the metasurface layer, and forming a mapping relationship with the corresponding structure of the metasurface.

[0083] According to the weight distribution obtained after weight training of the metasurface optical network, the accuracy test of the metasurface optical network is performed, and a first test result is obtained.

[0084] According to some embodiments, the accuracy test of the metasurface optical network is performed using the basic graphic data set as input data, and the obtained confusion matrix is ​​as follows: Figure 6 As shown. The horizontal axis is the predicted value of the input image by the set metasurface optical network, the vertical axis is the true value of the input image, and the numbers in the matrix represent the number. If the predicted value is consistent with the true value, it means that the classification of this figure is correct, this figure is assigned to the position corresponding to the predicted figure, and the value of this position is increased by 1. If the predicted value is inconsistent with the true value, it means that the classification of this figure is wrong, this figure is assigned to the position corresponding to the predicted figure outside the diagonal of the matrix, and the value of this position is increased by 1.

[0085] According to Figure 6 As shown in the confusion matrix, the accuracy test of the metasurface optical network corresponds to the first test result of 99.3%.

[0086] According to some embodiments, the pooling metasurface layer in the metasurface optical network can choose whether to train according to actual needs.

[0087] Furthermore, after the test of the metasurface optical network is completed, the weights of the electrical network are trained using a handwritten digital image dataset to obtain the weight distribution of the electrical network.

[0088] According to some embodiments, task switching is performed after the test of the metasurface optical network is completed, and the weight training of the electrical network is performed through a fully connected neural network with back propagation based on a handwritten digital image dataset.

[0089] According to the weight distribution obtained after the weight training of the electrical network, the accuracy test of the electrical network is performed, and a second test result is obtained.

[0090] According to some embodiments, the accuracy of the electrical network is tested using a handwritten digital image dataset as input data, and the obtained confusion matrix is ​​as follows: Figure 7As shown. The horizontal axis is the predicted value of the input image (0 to 9) by the set electrical network, the vertical axis is the true value of the input image (0 to 9), and the numbers in the matrix represent the number. If the predicted value is consistent with the true value, it means that the classification of this figure is correct, this figure is assigned to the position corresponding to the predicted figure, and the value of this position is increased by 1. If the predicted value is inconsistent with the true value, it means that the classification of this figure is wrong, this figure is assigned to the position corresponding to the predicted figure outside the diagonal of the matrix, and the value of this position is increased by 1.

[0091] According to Figure 7 As shown in the confusion matrix, the second test result corresponding to the accuracy test of the electrical network is 94.2%.

[0092] In step S400, when the test result meets the preset conditions, the optoelectronic computing chip is constructed through the network model.

[0093] When the accuracy of the first test result and the second test result both meet the preset conditions, the metasurface optical network is arrayed.

[0094] According to some embodiments, the preset condition can be set to an expected value of the accuracy test, such as 90%. That is, when the accuracy of the first test result and the second test result is greater than 90%, the array of the metasurface optical network is performed based on the set parameters to form at least one convolution metasurface layer array and at least one pooling metasurface layer array. Among them, at least one convolution metasurface layer array and at least one pooling metasurface layer array are used to form Figure 1 The universal diffraction calculation module in the optoelectronic computing chip shown.

[0095] According to some embodiments, when the first test result does not meet the preset conditions, the parameters of the metasurface optical network and the basic graphic data set used for training are adjusted, and the training is re-performed until the first test result meets the preset conditions.

[0096] Similarly, when the second test result does not meet the preset condition, the parameters of the electrical network and the handwritten digital image data set used for training are adjusted, and the training is performed again until the second test result meets the preset condition.

[0097] According to some embodiments, the weight of the electrical network can be adjusted arbitrarily according to the task target. That is, the weight of the electrical network can be re-determined according to the target to be identified to meet different target identification needs.

[0098] According to some embodiments, after the convolution metasurface layer array and the pooling metasurface layer array in the general diffraction calculation module are constructed, an electrical network is constructed based on the set parameters to form the following Figure 1 The image sensing module in the optoelectronic computing chip shown.

[0099] Figure 8 A flowchart of an image intelligent recognition method according to an exemplary embodiment of the present application is shown.

[0100] like Figure 8 As shown, in step S500, an optical signal containing image information is acquired.

[0101] In step S600, the optical signal is processed by the optoelectronic computing chip, and the processing result is obtained.

[0102] According to some embodiments, by Figure 3 The universal diffraction calculation module in the optoelectronic calculation chip constructed by the method shown extracts and compresses the features of the image information contained in the optical signal, and converts the optical signal that has been feature extracted and compressed into an electrical signal.

[0103] In step S700, the target to be identified in the image information is determined according to the processing result.

[0104] According to some embodiments, by Figure 3 The image sensing module in the optoelectronic computing chip constructed by the method shown processes the electrical signal containing the image information that has been feature extracted and compressed, thereby determining the target to be identified in the image information.

[0105] The embodiments of the present application are described in detail above, and the description of the above embodiments is only used to help understand the method and core idea of ​​the present application. At the same time, changes or deformations made by those skilled in the art based on the idea of ​​the present application, the specific implementation method and the scope of application of the present application, all belong to the scope of protection of the present application. In summary, the content of this specification should not be construed as limiting the present application.

Claims

1. A method for constructing an optoelectronic computing chip based on a metasurface, characterized in that: include: Get the preset training data set; Constructing a network model of an optoelectronic computing chip, wherein the network model includes a metasurface optical network and an electrical network; Training and testing the network model using the preset training data set, and obtaining test results; When the test result meets the preset conditions, the optoelectronic computing chip is constructed through the network model.

2. The method according to claim 1, characterized in that Construct a network model of an optoelectronic computing chip, including: Performing metasurface simulation modeling of the metasurface optical network; Setting parameters of the metasurface optical network; and Set the parameters of the electrical network.

3. The method according to claim 2, characterized in that Performing metasurface simulation modeling of the metasurface optical network includes: Constructing a hypersurface of the hypersurface optical network by a preset structure, wherein the hypersurface of the hypersurface optical network includes a preset number of neurons; The metasurface of the metasurface optical network is scanned by a preset wavelength to obtain the corresponding relationship between the preset number of neurons and the phase.

4. The method according to claim 2, characterized in that: The metasurface optical network includes at least one convolutional metasurface layer and at least one pooling metasurface layer; Setting parameters of the metasurface optical network includes: Configuring the at least one convolutional hypersurface layer and the at least one pooling hypersurface layer by a preset hypersurface layout and a preset array scale; The at least one pooling supersurface layer is configured by a preset distribution rule.

5. The method according to claim 2, characterized in that: Setting the parameters of the electrical network includes: The activation function of the electrical network is configured according to the category corresponding to the preset training data set for image classification and recognition.

6. The method according to claim 1, characterized in that The network model is trained and tested using the preset training data set, and the test results are obtained, including: Performing weight training on the metasurface optical network using the preset training data set to obtain a weight distribution of the metasurface optical network; An accuracy test of the metasurface optical network is performed according to the weight distribution of the metasurface optical network, and a first test result is obtained.

7. The method according to claim 1, characterized in that The network model is trained and tested using the preset training data set, and the test results are obtained, further comprising: Performing weight training on the electrical network using the preset training data set to obtain a weight distribution of the electrical network; An accuracy test of the electrical network is performed according to the weight distribution of the electrical network, and a second test result is obtained.

8. The method according to claim 6 or 7, characterized in that: When the test result meets the preset condition, constructing the optoelectronic computing chip through the network model includes: When the accuracy of the first test result and the second test result both meet the preset conditions, arraying the metasurface optical network to form at least one convolution metasurface layer array and at least one pooling metasurface layer array; The electrical network is constructed to form the optoelectronic computing chip with the at least one convolutional metasurface layer array and the at least one pooling metasurface layer array.

9. An optoelectronic computing chip, characterized in that: include: A general diffraction calculation module, comprising at least one convolution metasurface layer array and at least one pooling metasurface layer array, for extracting features of image information contained in a received optical signal through the at least one convolution metasurface layer array, and for compressing features of the image information for which features have been extracted through the at least one pooling metasurface layer array; An image sensing module, comprising an image sensor and an electrical network in a network model of an optoelectronic computing chip constructed by the method according to any one of claims 1 to 8, for converting an optical signal that has undergone feature extraction and feature compression of the image information into an electrical signal through the image sensor, and for classifying and identifying image information in the electrical signal through the electrical network; Wherein, the at least one convolutional super surface layer array and the at least one pooling super surface layer array are formed by the method as claimed in claim 8, and the at least one convolutional super surface layer array and the at least one pooling super surface layer array correspond to each other.

10. An image intelligent recognition method, characterized in that: include: Acquiring an optical signal containing image information; Processing the optical signal by the optoelectronic computing chip according to claim 9 and obtaining a processing result; According to the processing result, the target to be identified in the image information is determined.

Images