Optical element vectorization method, device, equipment and medium
By training optical components using a cascaded multi-layer perceptron network vectorization method, the problems of uncontrollable dimensions and loss of numerical information in traditional optical design are solved, and efficient automated design of optical systems is achieved.
Patent Information
- Application Number
- CN202511065476.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-31
- Publication Date
- 2025-09-30
- Estimated Expiration
- 2045-07-31
AI Technical Summary
Traditional optical design methods are difficult to meet the needs of high-dimensional optimization. There are problems such as uncontrollable dimensions and loss of numerical information, which leads to the optical system design relying on manual experience, long design cycle and high cost.
A cascaded multi-layer perceptron network is used to train the optical element vectorization method. By initializing the element vector configuration and back-propagation optimization, a fixed-dimensional vector representation of the optical element is generated. It is compatible with any number of parameters and supports optical path topology relationship analysis and system-level optimization.
It achieves dimensional uniformity and numerical fidelity of optical element vectors, supports optical system design based on graph neural networks, reduces manual feature engineering, and improves the degree of design automation.
Smart Images
Figure CN120577963B_ABST
Abstract
Description
Technical Field
[0001] The present invention mainly relates to the technical field of optical system vectorization characterization, and in particular to an optical element vectorization method, device, equipment and medium. Background Art
[0002] In the field of optics, with the exponential growth in the complexity of new optical systems such as metasurfaces and photonic integrated circuits, traditional design methods based on empirical formulas and parameter scanning have been unable to meet the needs of high-dimensional optimization. The rapid development of AI has driven the continuous transformation and reconstruction of traditional optical design methods. However, current AI-based optical design is limited to optical component design (such as optical component reverse design, optical imaging and detection, and photonic structure design) and system optimization levels (such as optical communication system optimization and photonic integrated circuit optimization). The system-level characterization is still at the stage of isolated parameter calibration and lacks coding technology that supports AI understanding. As a result, optical system design has problems such as reliance on manual experience, long design cycles, and high costs. To achieve comprehensive and accurate characterization of optical systems, vectorized representation of optical components is a key prerequisite. Traditional vectorized characterization methods have many limitations when dealing with optical component vectorization problems, such as:
[0003] Dimensionality uncontrollable problem: Traditional methods (such as One-hot encoding) lead to inconsistent vector lengths due to the difference in the number of parameters of different components. Figure 1 Figure 1 shows a diagram summarizing all optical elements used in an optical system in one embodiment. A plane mirror has two parameters: reflection wavelength range and reflectivity, while a concave mirror has three parameters: reflection wavelength range, focal length, and reflectivity. This makes it impossible to generate a vector representation of uniform dimensions, hindering batch processing of deep learning models.
[0004] Numerical information loss: Document embedding methods (such as Doc2vec) can fix the output dimension but cannot encode the specific numerical characteristics of the parameters. For example, lenses with focal lengths of 20mm and 50mm may be assigned similar vectors, resulting in distorted optical properties.
[0005] These limitations severely limit the ability of models like graph neural networks (GNNs) to analyze optical path topology, leading to a continued reliance on manual experience in optical system design, resulting in long design cycles and high costs. Therefore, a method for vectorizing optical components that simultaneously maintains dimensional uniformity and numerical fidelity is urgently needed to provide underlying support for intelligent optical system design. Summary of the Invention
[0006] In view of the technical problems existing in the prior art, the present invention proposes a method, device, equipment and medium for vectorizing an optical element.
[0007] To achieve the above object, the technical solution adopted by the present invention is as follows:
[0008] In one aspect, the present invention provides a method for vectorizing an optical element, comprising:
[0009] Obtaining different optical elements and their corresponding real optical parameters, based on which the major categories, minor categories, and functional characteristics of the different optical elements can be characterized, wherein the optical parameters include geometric parameters and material properties of the optical elements;
[0010] Based on each optical element and its corresponding optical parameters, an initialization element vector of the same length is configured for each optical element, wherein the first digit in the initialization element vector represents the major category of the optical element, the second digit represents the minor category of the optical element, and the third digit and each digit thereafter represents an optical parameter of the optical element;
[0011] Training a cascaded multilayer perceptron network based on initialization element vectors of each optical element and corresponding optical parameters;
[0012] Based on the trained cascade multi-layer perceptron network, the initialization element vector of the optical element to be predicted is optimized by back propagation until the vector convergence condition is met, and the final optical element vector representation is generated.
[0013] In another aspect, the present invention provides an optical element vectorization device, comprising:
[0014] The first module is used to obtain different optical elements and their corresponding real optical parameters, based on which the major categories, minor categories and functional characteristics of different optical elements can be characterized. The optical parameters include geometric parameters and material properties of the optical elements;
[0015] The second module is configured to configure an initialization element vector of the same length for each optical element based on each optical element and its corresponding optical parameter, wherein the first digit in the initialization element vector represents the major category of the optical element, the second digit represents the minor category of the optical element, and the third digit and each digit thereafter represents an optical parameter of the optical element;
[0016] The third module is used to train a cascaded multi-layer perceptron network based on the initialization element vectors of each optical element and the corresponding optical parameters;
[0017] The fourth module is used to optimize the initialization element vector of the optical element to be predicted by backpropagation based on the trained cascade multi-layer perceptron network until the vector convergence condition is met, thereby generating the final optical element vector representation.
[0018] Compared with the prior art, the present invention has the following beneficial effects:
[0019] To simultaneously meet the requirements of dimensional uniformity and numerical fidelity for optical component vectorization, the present invention provides an optical component vectorization method. Based on the individual optical components and their corresponding optical parameters, each optical component is initialized with an element vector of the same length. This method addresses the uncontrollable dimensionality of vectors with varying optical parameters. The method outputs fixed-dimensional vectors (e.g., 32 dimensions) compatible with optical components with any number of parameters. By constraining parameter prediction tasks, the method achieves high-fidelity encoding of key optical parameters, addressing the issue of numerical information loss. The generated optical component vectors can be directly input into a graph neural network (GNN), supporting optical path topology analysis and system-level optimization. This reduces manual feature engineering and improves the automation of optical system design. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in 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 only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.
[0021] Figure 1 This is a diagram summarizing all optical elements used in the optical system in one embodiment;
[0022] Figure 2 is a flow chart of a method for vectorizing an optical element provided by an embodiment;
[0023] Figure 3 A graph showing the cosine distance results between vectors of each sub-category of optical elements calculated in one embodiment;
[0024] Figure 4 This is an example of an embodiment of reducing the optical element vector to a 2D result graph using t-SNE;
[0025] Figure 5 This is an example of an embodiment where t-SNE is used to reduce the vector of the light source to a 2D result graph;
[0026] Figure 6 A schematic diagram of position distribution versus wavelength after reducing the LED vector to two dimensions using t-SNE in one embodiment;
[0027] Figure 7 Schematic diagram of position distribution versus power variation after t-SNE is used to reduce the LED vector to two dimensions in one embodiment. DETAILED DESCRIPTION
[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.
[0029] In one embodiment, referring to Figure 2 , providing an optical element vectorization method, comprising:
[0030] Obtaining different optical elements and their corresponding real optical parameters, based on which the major categories, minor categories, and functional characteristics of the different optical elements can be characterized, wherein the optical parameters include geometric parameters and material properties of the optical elements;
[0031] Based on each optical element and its corresponding optical parameters, an initialization element vector of the same length is configured for each optical element, wherein the first digit in the initialization element vector represents the major category of the optical element, the second digit represents the minor category of the optical element, and the third digit and each digit thereafter represents an optical parameter of the optical element;
[0032] Training a cascaded multilayer perceptron network based on initialization element vectors of each optical element and corresponding optical parameters;
[0033] Based on the trained cascade multi-layer perceptron network, the initialization element vector of the optical element to be predicted is optimized by back propagation until the vector convergence condition is met, and the final optical element vector representation is generated.
[0034] It can be understood that the major categories of optical components are light sources, lenses, prisms, polarizing elements, reflectors, filters, beam splitters, gratings, light detectors, light modulators, etc. The subcategories of optical components are the subcategories of the corresponding major categories, such as lenses are divided into concave lenses and convex lenses, etc. Figure 1 As shown. Optical parameters refer to the optical parameters of each optical element, including the geometric parameters (shape, size, dimensions of each dimension), material properties, performance parameters, etc. of the optical element.
[0035] Based on each optical element and its corresponding optical parameters, configure an initialization element vector of the same length for each optical element. For example, when initializing several optical elements, configure a trainable initialization element vector of length D for each optical element (D is a hyperparameter, for example, 32). For example, a convex lens with a focal length of 100mm has a label of [1,1,100], where the first digit represents the major category, the second represents the minor category, and the third and subsequent digits represent the focal length. Each element in the initialization element vector is uniformly randomly generated in the range [0,1]. For example, the 32-dimensional initialization element vector for a particular optical element is [0.21, -0.48,..., 0.37].
[0036] A cascaded multilayer perceptron network is constructed based on the statistically determined number and types (including major and minor categories) of optical elements and their optical parameters. The cascaded multilayer perceptron network is composed of a first multilayer perceptron, a second multilayer perceptron, and a third multilayer perceptron. The first multilayer perceptron determines the major category of the optical element (e.g., light source, lens, etc.) based on the initialization element vector. The second multilayer perceptron determines the minor category of the optical element (e.g., LED light source, laser light source, etc., or concave lens, convex lens, etc.) based on the major category determination result. The third multilayer perceptron predicts the optical parameters of the optical element (e.g., wavelength range and power of the LED light source, etc.) based on the minor category determination result. A loss function is constructed based on the prediction results of the first, second, and third multilayer perceptrons and the corresponding true values. With the goal of minimizing the loss function, the parameters of all multilayer perceptrons and the initialization element vector are updated using a backpropagation algorithm. The process is iterated continuously until a training stop condition is met (the training stop condition is either reaching a preset number of training rounds or the loss function converges to a set threshold). The parameters of all multilayer perceptrons are then fixed to obtain a trained cascaded multilayer perceptron network. Through the cascade MLP framework of optical component major category → optical component minor category → optical parameters, the hierarchical deconstruction of optical component characteristics is realized, solving the problem of unified characterization of heterogeneous components.
[0037] A loss function is constructed based on the prediction results of the first multilayer perceptron, the second multilayer perceptron, and the third multilayer perceptron and the corresponding true values. The loss function is:
[0038]
[0039] in 、 and They are 、 and The weight parameter, is the cross entropy loss between the category of optical elements judged by the first multi-layer perceptron and the real category of optical elements, is the cross entropy loss between the subclass of the optical element judged by the second multi-layer perceptron and the true subclass of the optical element, It is the mean square error loss or cross entropy loss of the optical parameters predicted by the third multi-layer perceptron relative to the true optical parameters.
[0040] The training stopping conditions for the cascaded multi-layer perceptron network can be reaching a preset number of training rounds (such as 500 rounds) or the loss function converges to a set threshold (such as the loss function decreases by <1e-5 for 10 consecutive rounds). It can also be set to stop when any of the following conditions are met: a. Reaching the maximum number of iterations (such as 500 rounds); b. The major category prediction accuracy is >0.99, the minor category prediction accuracy is >0.99, and the parameter prediction accuracy is >0.99.
[0041] Assume that the length of the initialization element vector is 32, the number of hidden nodes is 64, and there are three levels of multilayer perceptron layers in total, and each multilayer perceptron layer contains several multilayer perceptron models.
[0042] The first multilayer perceptron includes a multilayer perceptron model, and its model parameters are:
[0043] 32 nodes → 64 nodes → 9 nodes (softmax)
[0044] The 9 nodes in the third layer correspond to major categories of optical components such as light sources and lenses.
[0045] The second multilayer perceptron includes 9 multilayer perceptron models, and its model parameters are:
[0046] 32 nodes → 64 nodes → n nodes (softmax)
[0047] The value of n is related to the number of subcategories corresponding to each major category. For example, the value of n in the first multi-layer perceptron model is 3, which corresponds to the three subcategories of light sources: LED, laser, and ultrafast laser.
[0048] The third multilayer perceptron includes m multilayer perceptron models, and the value of m is related to the number of specific components with parameters. Figure 1 For example, here m is 23, and its model parameters are:
[0049] 32 nodes → 64 nodes → k nodes
[0050] The value of k is related to the number of parameters. For example, the k value of the first multi-layer perceptron is 2, which corresponds to the wavelength and power of the LED light source.
[0051] In the prediction stage, based on the trained cascaded multi-layer perceptron network, the initialization element vector of the optical element to be predicted is optimized through backpropagation until the vector convergence condition is met. The vector convergence condition is: reaching the preset number of vector update iterations or the L2 norm of the vector change is less than the set stability threshold for N consecutive times, generating the final optical element vector representation.
[0052] In the training phase, the present invention uses back propagation to simultaneously optimize model parameters and vector representation. In the prediction phase, the trained cascade multi-layer perceptron network parameters are fixed. Back propagation is specifically used for fine-tuning the vectors to improve the characterization quality of the vector representation of optical elements.
[0053] In a specific embodiment, during the training phase:
[0054] Initialize the element vector configuration: sort out all the optical elements used in the optical system, such as Figure 1 As shown, the broad categories are light sources, lenses, prisms, polarizers, reflectors, filters, beam splitters, gratings, photodetectors, and light modulators. For each optical component, 128 samples are randomly generated. Each component is assigned a trainable initialization vector of length 16. The initial values of each element in the initialization vector are uniformly randomly generated in the range [-1, 1]. For example, a convex lens with a focal length of 10 cm has the labels [1, 0, 10], where 1 represents the broad category, 0 represents the subcategory, and 10 represents the functional parameter. The initialization vector is: [0.21, -0.48, ..., 0.37] (16 dimensions).
[0055] Construct a cascaded multi-layer perceptron network:
[0056] The first multilayer perceptron for broad category judgment: Input layer (16 neurons) → Hidden layer (64 neurons, ReLU) → Output layer (K neurons, softmax). Operation: Input the initialization element vector and output the broad category probability distribution. For example, the output probability of a lens is 0.85, the probability of a prism is 0.01, and so on.
[0057] Second multilayer perceptron for small-category classification: Input layer (16 neurons) → Hidden layer (48 neurons, ReLU) → Output layer (M neurons, softmax). Operation: For components whose large category is a lens, output the probability distribution of the small category (e.g., convex lens: 0.82, concave lens: 0.18).
[0058] The third multilayer perceptron for optical parameter prediction: Input layer (16 neurons) → Hidden layer (32 neurons, ReLU) → Output layer (P neurons, linear activation). Operation: For the convex lens sub-class, output parameter predictions (e.g., focal length = 10 cm).
[0059] The overall loss function is constructed as:
[0060]
[0061] in 、 and They are 、 and The weight parameter, is the cross entropy loss between the category of optical elements judged by the first multi-layer perceptron and the real category of optical elements, is the cross entropy loss between the subclass of the optical element judged by the second multi-layer perceptron and the true subclass of the optical element, It is the mean square error loss or cross entropy loss of the optical parameters predicted by the third multi-layer perceptron relative to the true optical parameters.
[0062] Backpropagation update object: Update the weight parameters and initialization element vectors of all multilayer perceptrons at the same time, with a learning rate of 0.001.
[0063] Cascade multi-layer perceptron network training termination judgment: stop when any of the following conditions are met: a. reaching the maximum number of iterations (such as 500 rounds); b. the large category prediction accuracy > 0.99, the small category prediction accuracy > 0.99, the parameter prediction R 2 >0.99. Fix all weight parameters of the trained cascaded multilayer perceptron network.
[0064] Prediction phase: Initialize the label of the optical element to be predicted (e.g., [1,1,100] for a convex lens) and the initial component vector ([0.21, -0.48, ..., 0.37] (16-dimensional)). Based on the trained cascaded multilayer perceptron network, backpropagation is used to optimize the initial component vector of the optical element to be predicted until the vector convergence condition is met, generating the final optical component vector representation. After completing these steps, the final optical component vector is obtained by repeatedly updating the initial component vector.
[0065] The cosine distance between the vectors of each sub-category of optical elements calculated in this embodiment is as follows: Figure 3 At the same time, the optical element vector is reduced to 2 dimensions using t-SNE (t-Distributed Stochastic Neighbor Embedding), and the result is as follows Figure 4 As shown, from Figure 4 It can be seen that the vectors of optical elements belonging to the same category are closer. Figure 4 Extract the results of the light source category, such as Figure 5As shown in Figure 1, the result of using t-SNE to reduce the vector of the light source to 2D is shown. It can be seen that the optical element vector can also reflect the specific sub-category information of the optical element. Figure 6 This is a diagram showing how the position distribution of LEDs changes with wavelength after t-SNE is used to reduce their vectors to two dimensions. Figure 7 This figure shows the position distribution of the LED after the vector is reduced to two dimensions using t-SNE. The size of the circle represents the numerical information of wavelength and power. It can be seen that the optical element vector can simultaneously reflect the different functional attribute parameters of the optical element.
[0066] Another embodiment provides an optical element vectorization device, comprising:
[0067] The first module is used to obtain different optical elements and their corresponding real optical parameters, based on which the major categories, minor categories and functional characteristics of different optical elements can be characterized. The optical parameters include geometric parameters and material properties of the optical elements;
[0068] The second module is configured to configure an initialization element vector of the same length for each optical element based on each optical element and its corresponding optical parameter, wherein the first digit in the initialization element vector represents the major category of the optical element, the second digit represents the minor category of the optical element, and the third digit and each digit thereafter represents an optical parameter of the optical element;
[0069] The third module is used to train a cascaded multi-layer perceptron network based on the initialization element vectors of each optical element and the corresponding optical parameters;
[0070] The fourth module is used to optimize the initialization element vector of the optical element to be predicted by backpropagation based on the trained cascade multi-layer perceptron network until the vector convergence condition is met, thereby generating the final optical element vector representation.
[0071] On the other hand, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the optical element vectorization method provided in any of the above embodiments are implemented. The computer device may be a server. The computer device comprises a processor, a memory, a network interface, and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store sample data. The network interface of the computer device is used to communicate with an external terminal via a network connection.
[0072] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the optical element vectorization method provided in any of the above embodiments are implemented.
[0073] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0074] Matters not covered by the present invention are known technologies.
[0075] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0076] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the spirit of the present application, and such modifications and improvements are all within the scope of protection of the present application.
[0077] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.
Claims
1. An optical element vectorization method, characterized in that: include: Obtaining different optical elements and their corresponding real optical parameters, based on which the major categories, minor categories, and functional characteristics of the different optical elements can be characterized, wherein the optical parameters include geometric parameters and material properties of the optical elements; Based on each optical element and its corresponding optical parameters, an initialization element vector of the same length is configured for each optical element, wherein the first digit in the initialization element vector represents the major category of the optical element, the second digit represents the minor category of the optical element, and the third digit and each digit thereafter represents an optical parameter of the optical element; A cascaded multilayer perceptron network is trained based on the initialization element vectors of each optical element and the corresponding optical parameters, wherein the cascaded multilayer perceptron network is formed by cascading a first multilayer perceptron, a second multilayer perceptron, and a third multilayer perceptron; the first multilayer perceptron is used to judge the major category of the optical element according to the initialization element vector; the second multilayer perceptron is used to judge the minor category of the optical element according to the major category judgment result; the third multilayer perceptron is used to predict the optical parameters of the optical element according to the minor category judgment result; a loss function is constructed based on the prediction results of the first multilayer perceptron, the second multilayer perceptron, and the third multilayer perceptron and the corresponding true values, and with the minimum loss function as the goal, the parameters of all multilayer perceptrons and the initialization element vector are updated through a back propagation algorithm, and the iteration is continued until the training stop condition is met, and the parameters of all multilayer perceptrons are fixed to obtain a trained cascaded multilayer perceptron network; Based on the trained cascade multi-layer perceptron network, the initialization element vector of the optical element to be predicted is optimized by back propagation until the vector convergence condition is met, and the final optical element vector representation is generated.
2. The optical element vectorization method according to claim 1, characterized in that: The loss function is: in 、 and They are 、 and The weight parameter, is the cross entropy loss between the category of optical elements judged by the first multi-layer perceptron and the real category of optical elements, is the cross entropy loss between the subclass of the optical element judged by the second multi-layer perceptron and the true subclass of the optical element, It is the mean square error loss or cross entropy loss of the optical parameters predicted by the third multi-layer perceptron relative to the true optical parameters.
3. The optical element vectorization method according to claim 1, characterized in that: The training stop condition is that the preset number of training rounds is reached or the loss function converges to the set threshold.
4. The optical element vectorization method according to any one of claims 1 to 3, characterized in that: The initialization element vector of the optical element to be predicted is optimized until the vector convergence condition is met, where the vector convergence condition is: reaching a preset number of vector update iterations or the L2 norm of the vector change is less than a set stability threshold for N consecutive times.
5. An optical element vectorization device, characterized in that: include: The first module is used to obtain different optical elements and their corresponding real optical parameters, based on which the major categories, minor categories and functional characteristics of different optical elements can be characterized. The optical parameters include geometric parameters and material properties of the optical elements. The second module is configured to configure an initialization element vector of the same length for each optical element based on each optical element and its corresponding optical parameter, wherein the first digit in the initialization element vector represents the major category of the optical element, the second digit represents the minor category of the optical element, and the third digit and each digit thereafter represents an optical parameter of the optical element; The third module is used to train a cascaded multi-layer perceptron network based on each optical element and the initialization element vector of the corresponding optical parameter, wherein the cascaded multi-layer perceptron network is formed by cascading a first multi-layer perceptron, a second multi-layer perceptron, and a third multi-layer perceptron; using the first multi-layer perceptron to judge the major category of the optical element according to the initialization element vector; using the second multi-layer perceptron to judge the minor category of the optical element according to the major category judgment result; using the third multi-layer perceptron to predict the optical parameters of the optical element according to the minor category judgment result; constructing a loss function based on the prediction results of the first multi-layer perceptron, the second multi-layer perceptron, and the third multi-layer perceptron and the corresponding true values, and updating the parameters of all multi-layer perceptrons and the initialization element vector by a back propagation algorithm with the goal of minimizing the loss function, and continuously iterating until the training stop condition is met, fixing the parameters of all multi-layer perceptrons, and obtaining a trained cascaded multi-layer perceptron network; The fourth module is used to optimize the initialization element vector of the optical element to be predicted by backpropagation based on the trained cascade multi-layer perceptron network until the vector convergence condition is met, thereby generating the final optical element vector representation.
6. The optical element vectorization device according to claim 5, characterized in that: The loss function constructed in the third module is: in 、 and They are 、 and The weight parameter, is the cross entropy loss between the category of optical elements judged by the first multi-layer perceptron and the real category of optical elements, is the cross entropy loss between the subclass of the optical element judged by the second multi-layer perceptron and the true subclass of the optical element, It is the mean square error loss or cross entropy loss of the optical parameters predicted by the third multi-layer perceptron relative to the true optical parameters.
7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the optical element vectorization method as claimed in claim 1 are implemented.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the optical element vectorization method according to claim 1 are implemented.