A method for extracting performance change trend of optical lens based on gradient analysis
By constructing an optical lens performance prediction model based on gradient analysis and utilizing neural networks and tree-structured fully connected networks, the problem of relying on expert experience in optical lens design is solved, enabling rapid and accurate extraction of lens performance change trends and improving design efficiency and quality.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2023-07-28
- Publication Date
- 2026-06-19
AI Technical Summary
Existing optical lens design methods rely on expert experience and lack methods for extracting performance change trends based on gradient analysis, resulting in low design efficiency and difficulty in guaranteeing quality.
By using gradient analysis, an optical lens performance prediction model is constructed. Using neural networks and tree-structured fully connected networks, the performance variation trend of lens parameters is extracted. This includes preprocessing simulation data, constructing a training set, iteratively training the model, and calculating derivatives, thereby achieving rapid and accurate extraction of performance indicators.
It improves the accuracy of lens performance index regression, shortens design time, enhances design quality, provides low-cost design assistance, and helps to quickly locate reasonable design solutions.
Smart Images
Figure CN116974066B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of performance change trend extraction in the optical lens design process, and particularly to a method for extracting optical lens performance change trends based on gradient analysis. Background Technology
[0002] With the rapid development of optical lens design technology, computer-aided design techniques based on deep learning and machine learning are becoming increasingly mature. Among these techniques, the performance trends of optical lenses play a crucial role in the design process.
[0003] The mainstream approach to optical lens design still relies on expert experience, with the final solution determined manually from overall design to detailed specifications. However, several solutions based on heuristic or machine learning algorithms have emerged in the past. The traditional approach involves defining multiple variables and an objective function, combining global optimization algorithms and computer-aided design (CAD) tools to find numerous possible design solutions; then, the optical lens designer verifies the usability of one or more optical lens parameter configurations. End-to-end algorithms can directly search for available optical lens designs using heuristic algorithms such as genetic algorithms, simulating several solutions to solve the optimization problem.
[0004] While the aforementioned solutions apply machine learning algorithms to industrial design and tolerance design, most research, as described above, focuses solely on end-to-end problem-solving. However, some R&D companies are now adopting a more balanced approach, using low-cost simulation data as a guide on top of manual design to create a traceable and interpretable design and manufacturing workflow. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for extracting the performance variation trend of optical lenses based on gradient analysis. This invention can extract the performance variation trend of parameter pairs under a single optical lens design scheme.
[0006] The objective of this invention is achieved through the following technical solution: a method for extracting the performance change trend of optical lenses based on gradient analysis, comprising the following steps:
[0007] (1) Determine the value constraints based on the preset yield rate and the physical range of each lens parameter, and sample the lens parameters to generate a simulation dataset; wherein, each sample in the simulation dataset contains the same lens parameters as the actual designed lens and the corresponding simulation results, and the simulation results are modulation transfer functions;
[0008] (2) Preprocess the simulation dataset to obtain lens performance-related values as the task objective;
[0009] (3) Obtain the lens parameters of multiple lenses in the lens according to the preprocessed simulation dataset, stack the lens parameters according to the assembly order of multiple lenses in the lens to obtain the lens parameters; construct a training set according to the lens parameters and the task objective;
[0010] (4) Construct an optical lens performance prediction model based on a neural network, use the training set to iteratively train the optical lens performance prediction model, and adjust the parameters of the optical lens performance prediction model according to the loss function to obtain a well-trained optical lens performance prediction model.
[0011] (5) Input the lens parameters of the optical lens into the trained optical lens performance prediction model to obtain the regression value of the lens performance index.
[0012] (6) Calculate the derivative of the regression value of each lens performance index with respect to the lens parameters to form the trend of the lens parameters with respect to the lens performance index.
[0013] Further, step (1) specifically includes: setting a preset yield rate, defining the physical range of each lens parameter, uniformly sampling lens parameters to generate simulation samples under the constraint of the physical range of lens parameters, calculating the yield rate based on the simulation results, and determining whether the calculated yield rate is less than the preset yield rate. If the calculated yield rate is less than the preset yield rate, the physical range of lens parameters is reduced; otherwise, the physical range of lens parameters is increased. This process samples lens parameters and their simulation results to generate a simulation dataset.
[0014] Furthermore, step (2) specifically includes: deleting values in the sampled values of the modulation transfer function that are irrelevant to the lens performance, so as to obtain lens performance-related values as the task objective.
[0015] Furthermore, step (4) includes the following sub-steps:
[0016] (4.1) An optical lens performance prediction model is constructed based on a neural network, the optical lens performance prediction model including a feature extractor and a tree-structured fully connected network;
[0017] (4.2) Input the lens parameters in the training set into the optical lens performance prediction model to obtain the lens performance index regression value. Calculate the loss function based on the lens performance index regression value and the task objective in the training set. Use the loss function to use the optimizer to update the parameters of the optical lens performance prediction model in reverse to obtain the trained optical lens performance prediction model.
[0018] Furthermore, the feature extractor includes a first feature extraction module, a second feature extraction module, and a third feature extraction module connected in sequence. Each of the first, second, and third feature extraction modules includes a convolutional layer, an activation function layer, and a max pooling layer. The convolutional layer, activation function layer, and max pooling layer are connected in sequence. The convolutional layer of the first feature extraction module is connected to the convolutional layer of the second feature extraction module via a bypass connection. The max pooling layer of the first feature extraction module is connected to the max pooling layer of the second feature extraction module via a bypass connection. The convolutional layer of the second feature extraction module is connected to the convolutional layer of the third feature extraction module via a bypass connection. The max pooling layer of the second feature extraction module is connected to the max pooling layer of the third feature extraction module via a bypass connection.
[0019] Furthermore, the number of layers in the tree-shaped fully connected network is determined based on the inter-target correlation. The first fully connected layer is a single fully connected layer, and the number of fully connected layers in each of the remaining fully connected layers is determined based on the experimental conditions: it is determined whether there is an inter-target correlation among the target performance indicators generated within the same field of view. If there is an inter-target correlation, the number of layers in the tree-shaped fully connected network is set to three. The first fully connected layer is a single fully connected layer, and the number of fully connected layers in the second fully connected network is determined based on the number of fields of view in the experimental conditions. In the third fully connected network, each upper fully connected layer connects to K fully connected layers, where the value of K is determined based on the number of experiments in each field of view in the experimental conditions. If there is no inter-target correlation, the number of layers in the tree-shaped fully connected network is set to two. The first fully connected layer is a single fully connected layer, and the number of fully connected layers in the second fully connected network is determined based on the number of fields of view and the number of experiments in the experimental conditions.
[0020] Further, step (4.2) specifically involves: inputting the lens parameters from the training set into the optical lens performance prediction model to obtain the lens performance index regression value; calculating the loss function based on the lens performance index regression value and the task objective in the training set; using an optimizer to calculate the gradient of the loss function for each batch of samples; smoothing the gradient; and then backpropagating to update the parameters of the optical lens performance prediction model; repeating the training until the loss function is less than a preset loss threshold to obtain a trained optical lens performance prediction model.
[0021] Preferably, the optimizer is the Adam optimizer.
[0022] Furthermore, the loss function uses mean squared error, and its expression is:
[0023]
[0024] Among them, y iIt is the regression value of the lens performance index for the i-th sample. is the task objective of the i-th sample, and n is the total number of samples in the training set.
[0025] Furthermore, the formula for calculating the derivative is:
[0026]
[0027] Among them, y i x is the regression value of the i-th lens performance index. j This is the j-th camera parameter, and the activation values of the k hidden layers are denoted as h1 to h2 respectively. k The required derivative I ij For y i For x j The partial derivative of y, when expanded, is y i Differentiate the activation values of the hidden layers layer by layer.
[0028] The beneficial effects of this invention are as follows: By introducing a tree network structure into the optical lens performance prediction model, this invention utilizes parameter sharing to analyze the correlation between targets, thereby improving the accuracy of performance index regression and enhancing the accuracy of trend extraction; After training the optical lens performance prediction model, this invention eliminates the need for initial solutions and repeated iterative calculations during the design phase, requiring only one forward prediction and one backward differentiation to extract the performance trend, significantly shortening design time, improving design efficiency, and greatly enhancing design quality; This invention enables rapid and accurate extraction of optical lens performance trends at low cost, providing design assistance to designers in the early stages of design and helping them to more quickly identify reasonable design solutions. Attached Figure Description
[0029] Figure 1 This is a flowchart of the method for extracting the performance change trend of optical lenses based on gradient analysis according to the present invention;
[0030] Figure 2 This is a schematic diagram illustrating the logic for obtaining lens parameters in this invention.
[0031] Figure 3 This is a schematic diagram of a structure of the optical lens performance prediction model in this invention;
[0032] Figure 4 This is another structural schematic diagram of the optical lens performance prediction model in this invention;
[0033] Figure 5 This is a schematic diagram of the network structure of the feature extractor in this invention;
[0034] Figure 6This is a schematic diagram of a visual image illustrating the trend of lens performance changes after dimensionality reduction, provided as an embodiment of the present invention. Detailed Implementation
[0035] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims. It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not intended to limit this application.
[0036] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.
[0037] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "in response to determination," or "includes." Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process or method. Without further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0038] To improve the efficiency and quality of optical lens design and production tolerance standard setting, this invention provides a method for extracting optical lens performance change trends based on gradient analysis, which will be described in detail below.
[0039] See Figure 1 The method for extracting the performance change trend of optical lenses based on gradient analysis of the present invention specifically includes the following steps:
[0040] (1) Determine the value constraints based on the preset yield rate and the physical range of each lens parameter, and generate a simulation dataset by sampling the lens parameters.
[0041] In this embodiment, value constraints are determined based on a preset yield rate and the physical range of each lens parameter. The lens parameters are then sampled to generate a simulation dataset. Specifically, this includes: setting a preset yield rate; defining the physical range of each lens parameter; uniformly sampling lens parameters within the constraints of the physical range to generate simulation samples; calculating the yield rate based on the simulation results; and determining whether the calculated yield rate is less than a preset yield rate. If the calculated yield rate is less than the preset yield rate, the physical range of the lens parameters is reduced; otherwise, the physical range of the lens parameters is increased. This process samples the lens parameters and their simulation results to generate the simulation dataset.
[0042] Specifically, the yield rate is first set to 50%, and then the physical range of each lens parameter is defined, imposing value constraints. Under the constraints of the physical range of the lens parameters, lens parameter values are uniformly sampled to generate simulation samples. Then, the yield rate is calculated based on the simulation results. If the yield rate does not meet the requirements, i.e., the calculated yield rate is less than the preset yield rate of 50%, the physical range of the lens parameters is narrowed; otherwise, the physical range of the lens parameters is widened. The simulation dataset is generated through the above process.
[0043] Furthermore, each sample in the simulation dataset contains the same lens parameters as the actual designed lens and the corresponding simulation results, which are the modulation transfer function (MTF). For example, if the lens parameters of the actual designed lens are x1, x2, and x3, and the sampled values of the MTF are mtf1 and mtf2, then the samples in the simulation dataset also contain the lens parameters x1, x2, and x3, and the sampled values of the MTF mtf1 and mtf2.
[0044] It should be understood that, since the simulation dataset contains lens parameters and their corresponding simulation results, which are the modulation transfer functions and contain performance-related values, the simulation dataset can be understood as a performance prediction simulation dataset for optical lenses.
[0045] (2) Preprocess the simulation dataset to obtain lens performance-related values as the task objective.
[0046] In this embodiment, the simulation dataset is preprocessed to obtain lens performance-related values. Specifically, this includes deleting sampled values of the modulation transfer function that are irrelevant to lens performance, so as to obtain lens performance-related values as the task objective.
[0047] It should be understood that the modulation transfer function (MTF) expresses how the modulation of an object is transmitted to the image through the lens. In actual experiments, the MTF result is a broken line containing multiple sampled values. These sampled values include all the values actually used to judge lens performance, i.e., lens performance-related values. Other values are removed, retaining only the lens performance-related values as the task objective.
[0048] It should be noted that different lens parameters will affect the shape of the resulting input data when preprocessing the simulation dataset. For example, if the lens has 7 lenses, each containing 8 parameters, the input data will be 7×8 two-dimensional data; if the lens has 6 lenses, each containing 9 parameters, the input data will be 6×9 two-dimensional data.
[0049] (3) Obtain the lens parameters of multiple lenses in the lens according to the preprocessed simulation dataset, stack the lens parameters according to the assembly order of multiple lenses in the lens to obtain the lens parameters; construct a training set according to the lens parameters and the task objective.
[0050] Among them, lens parameters are one-dimensional data, and lens parameters are two-dimensional data.
[0051] Specifically, such as Figure 2 As shown, the lens comprises multiple lens elements. Based on the preprocessed simulation dataset, the lens parameters of these elements can be obtained. Following the assembly order of the lens elements, the parameters of each lens element are stacked sequentially. Since the lens parameters are one-dimensional data, stacking them yields two-dimensional data, which represents the lens parameters. It should be understood that adjacent lens parameters indicate that the actual lens elements are also adjacent. Stacking the lens parameters according to the assembly order allows for better extraction of the interaction features between the lens elements, improving the accuracy of subsequent regression tasks.
[0052] (4) Construct an optical lens performance prediction model based on a neural network, use the training set to iteratively train the optical lens performance prediction model, and adjust the parameters of the optical lens performance prediction model according to the loss function to obtain a well-trained optical lens performance prediction model.
[0053] (4.1) An optical lens performance prediction model is constructed based on a neural network. The optical lens performance prediction model includes a feature extractor and a tree-structured fully connected network.
[0054] In this embodiment, the feature extractor includes a first feature extraction module, a second feature extraction module, and a third feature extraction module connected in sequence, such as... Figure 5As shown, the first, second, and third feature extraction modules each include a convolutional layer, an activation function layer, and a max-pooling layer. These layers are sequentially connected. The convolutional layers of the first and second feature extraction modules are connected via a shortcut (i.e., a residual mechanism). Similarly, the max-pooling layers of the first and second feature extraction modules are connected via a shortcut, as are the convolutional layers of the second and third feature extraction modules. By inputting lens parameters into the feature extractor, high-dimensional features can be extracted from the lens parameters. Through multi-layer processing, sufficiently high-dimensional features can be obtained.
[0055] Preferably, the size of the convolution kernel is 2×2, the pooling area of the max pooling layer is 2×2, and the activation function layer uses the Leaky ReLU activation function layer.
[0056] It should be understood that the introduction of Leaky ReLU greatly reduces the adverse effects of non-negligible negative inputs on feature extractor training in this task, and retains the good performance of ReLU to some extent; the introduction of the residual mechanism reduces the problem of gradient vanishing and training failure when the feature extractor is too deep. Both make the training of the feature extractor more efficient and stable, and improve the performance of the feature extractor.
[0057] In this embodiment, the number of layers in the tree-structured fully connected network is determined based on the correlation between the targets. The first fully connected layer is a single fully connected layer, and the number of fully connected layers in each of the remaining fully connected networks is determined based on the experimental conditions.
[0058] Specifically, the algorithm determines whether there is inter-target correlation among target performance indicators generated within the same field of view. If inter-target correlation exists, the tree-structured fully connected network is set to three layers: the first layer is a single fully connected layer; the number of fully connected layers in the second layer is determined based on the number of fields of view in the experimental conditions; and each upper fully connected layer in the third layer connects to K fully connected layers, where the value of K is determined based on the number of experiments in each field of view in the experimental conditions. If there is no inter-target correlation, the tree-structured fully connected network is set to two layers: the first layer is a single fully connected layer; and the number of fully connected layers in the second layer is determined based on the number of fields of view and the number of experiments in the experimental conditions. The high-dimensional features extracted by the feature extractor are input into the tree-structured fully connected network to obtain the regression values of the lens performance indicators, which correspond to the sampled values in the modulation transfer function used as lens performance indicators.
[0059] For example, two experiments are conducted in 16 fields of view to obtain 32 target performance indicators. Here, one field of view condition is used by two experiments. It is determined whether there is inter-target correlation among the target performance indicators generated within the same field of view. If there is no inter-target correlation, the tree-structured fully connected network should be set as a two-layer fully connected network. The first layer is a single fully connected layer. The number of fully connected layers in the second layer is determined based on the number of fields of view and the number of experiments in the experimental conditions, 16 × 2 = 32, meaning the second layer has 32 fully connected layers. Each fully connected layer in the first layer connects to 32 other fully connected layers, and these 32 fully connected layers constitute the second layer. Figure 4 As shown. If there is inter-target correlation among the target performance indicators generated within the same field of view, the tree-structured fully connected network should be set as a three-layer fully connected network. The first layer is a single fully connected layer. The number of fully connected layers in the second layer is determined according to the number of fields of view in the experimental conditions, i.e., a single fully connected layer in the first layer connects to 16 fully connected layers, and these 16 fully connected layers constitute the second layer. In the third layer, each upper fully connected layer connects to K fully connected layers, where the value of K is determined according to the number of experiments in each field of view in the experimental conditions, i.e., each fully connected layer in the second layer connects to 2 fully connected layers, for a total of 32 fully connected layers, and these 32 fully connected layers constitute the third layer. Figure 3 As shown.
[0060] For example, consider conducting three experiments in ten fields of view to obtain 30 target performance metrics. If there is no inter-target correlation among the target performance metrics generated within the same field of view, the structure of the tree-structured fully connected network is as follows: the first fully connected layer is a single fully connected layer, and this single fully connected layer connects to 30 other fully connected layers, forming the second fully connected layer. If there is inter-target correlation among the target performance metrics generated within the same field of view, the structure of the tree-structured fully connected network is as follows: the first fully connected layer is a single fully connected layer, and this single fully connected layer connects to 10 other fully connected layers, forming the second fully connected layer. Each fully connected layer in the second fully connected network connects to 3 other fully connected layers, for a total of 10 × 3 = 30 fully connected layers, forming the third fully connected layer.
[0061] In this embodiment, the optical lens performance prediction model includes a feature extractor and a tree-structured fully connected network. Lens parameters are input into the optical lens performance prediction model to obtain predicted regression values for lens performance indicators. Specifically, the lens parameters are first processed by the feature extractor to obtain high-dimensional features, which are then processed by the tree-structured fully connected network to obtain predicted regression values for lens performance indicators. The partial derivatives of each lens performance indicator regression value with respect to the lens parameters are then calculated to obtain the trend of the lens performance indicators. The tree-structured fully connected network in the optical lens performance prediction model can analyze targets that may have inter-field and intra-field correlations in stages, thereby fully analyzing the correlations between multiple targets and ultimately obtaining highly accurate lens performance indicator values.
[0062] (4.2) Input the lens parameters in the training set into the optical lens performance prediction model to obtain the lens performance index regression value. Calculate the loss function based on the lens performance index regression value and the task objective in the training set. Use the loss function to use the optimizer to update the parameters of the optical lens performance prediction model in reverse to obtain the trained optical lens performance prediction model.
[0063] In this embodiment, the training process of the optical lens performance prediction model is as follows: the lens parameters in the training set are input into the optical lens performance prediction model to obtain the regression value of the lens performance index; the loss function is calculated based on the regression value of the lens performance index and the task objective in the training set; the gradient of the loss function for each batch of samples is calculated using an optimizer, and the gradient is smoothed; then the parameters of the optical lens performance prediction model are updated by backpropagation; the training is repeated until the loss function is less than the preset loss threshold to obtain the trained optical lens performance prediction model.
[0064] Preferably, the optimizer is the Adam (Adaptive Moment Estimation) optimizer.
[0065] Specifically, training the optical lens performance prediction model is an iterative optimization process. Lens parameters from the training set are input into the optical lens performance prediction model to obtain regression values of lens performance indicators. A loss function is calculated based on the regression values and the task objective in the training set. The Adam optimizer is used to calculate the gradient of the loss function for each batch of samples, and the gradient is smoothed. Then, backpropagation is used to propagate the gradient back into the optical lens performance prediction model, updating the parameters of the model and training the entire model. This process gradually reduces the loss value of the network until optimal performance is achieved. Training stops when the loss function is less than a preset loss threshold. In this embodiment, the loss threshold is set to 0.000025; that is, training stops when the loss calculated according to the loss function is less than 0.000025, completing the training of the optical lens performance prediction model. Saving the latest parameters of the optical lens performance prediction model yields the trained optical lens performance prediction model.
[0066] Furthermore, the loss function uses the mean squared error (MSE), which is expressed as follows:
[0067]
[0068] Among them, y i It is the regression value of the lens performance index for the i-th sample. is the task objective of the i-th sample, and n is the total number of samples in the training set.
[0069] (5) Input the lens parameters of the optical lens into the trained optical lens performance prediction model to obtain the regression value of the lens performance index.
[0070] Specifically, the lens parameters of the actual optical lens are input into the trained optical lens performance prediction model to obtain regression values for lens performance indicators. The optical lens performance prediction model can be used to obtain regression values for various lens performance indicators.
[0071] It should be understood that since each fully connected layer in the last fully connected network corresponds to an output, there are N fully connected layers in the last fully connected network, which corresponds to N regression values for lens performance indicators.
[0072] (6) Calculate the derivative of the regression value of each lens performance index with respect to the lens parameters to form the trend of the lens parameters with respect to the lens performance index.
[0073] Furthermore, the formula for calculating the derivative is:
[0074]
[0075] Among them, y i x is the regression value of the i-th lens performance index. j This is the j-th camera parameter, and the activation values of the k hidden layers are denoted as h1 to h2 respectively. k The required derivative I ij For y i For x j The partial derivative of y, when expanded, is y i Differentiate the activation values of the hidden layers layer by layer.
[0076] For example, the derivatives of the calculated regression values of each lens performance index with respect to the lens parameters are shown in Table 1. The calculated derivatives can form a trend, thus obtaining the trend of the lens performance index. This trend can then be visualized, as shown in the visualization image below. Figure 6 As shown. The t-SNE method is used to reduce the dimensionality of the independent variables, and each dependent variable can be obtained as follows. Figure 6 The image shown is a visualization of a trend. From Figure 6 As can be seen, within a certain range, the extracted gradients and indicators show the same trend. This demonstrates that gradient analysis can be effectively used to extract the changing trends of lens performance indicators.
[0077] Table 1: Derivatives of regression values of various lens performance indicators with respect to lens parameters
[0078] <![CDATA[x1]]> <![CDATA[x2]]> <![CDATA[x3]]> … <![CDATA[x n ]]> <![CDATA[y1]]> <![CDATA[I 11 ]]> <![CDATA[I 12 ]]> <![CDATA[I 13 ]]> … <![CDATA[I 1n <!-- 7 -->]]> … … … … … … <![CDATA[y m ]]> <![CDATA[I m1 ]]> <![CDATA[I m2 ]]> <![CDATA[I m3 ]]> … <![CDATA[I mn ]]>
[0079] In summary, when designing optical lenses using the scheme provided in this invention, low-cost simulation data can be fully utilized to extract performance change trends for optical lens schemes with various parameter scales. This provides design assistance to designers in the early stages of design, helping them to quickly locate a reasonable design scheme. This trend can also be used to quickly identify tolerance standards suitable for production. Furthermore, the introduction of a tree network structure into the optical lens performance prediction model, utilizing parameter sharing to analyze the correlation between targets, improves the accuracy of performance index regression, and consequently enhances the accuracy of the extracted trends in subsequent steps.
[0080] Furthermore, after training the optical lens performance prediction model, this invention eliminates the need for initial solutions and repeated iterative calculations during the design phase. Only one forward prediction and one backward differentiation are required to extract the performance change trend, significantly reducing design time and greatly improving design quality. This invention enables rapid and accurate extraction of optical lens performance change trends at low cost.
[0081] Of course, implementing any product or method of the present invention does not necessarily require achieving all of the advantages described above at the same time.
[0082] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for extracting performance change trend of an optical lens based on gradient analysis, characterized in that, Includes the following steps: (1) Determine the value constraints based on the preset yield rate and the physical range of each lens parameter, and generate a simulation dataset by sampling the lens parameters; wherein, each sample in the simulation dataset contains the same lens parameters as the actual designed lens and the corresponding simulation results, and the simulation results are modulation transfer functions; (2) Preprocess the simulation dataset to obtain lens performance-related values as the task objective; (3) Obtain the lens parameters of multiple lenses in the lens according to the preprocessed simulation dataset, stack the lens parameters according to the assembly order of multiple lenses in the lens to obtain the lens parameters; construct a training set according to the lens parameters and the task objective; (4) Construct an optical lens performance prediction model based on a neural network. The optical lens performance prediction model includes a feature extractor and a tree-structured fully connected network. Iteratively train the optical lens performance prediction model using a training set and adjust the parameters of the optical lens performance prediction model according to the loss function to obtain a well-trained optical lens performance prediction model. The number of layers in the tree-shaped fully connected network is determined based on the correlation between targets. The first fully connected layer is a single fully connected layer, and the number of fully connected layers in each of the remaining fully connected layers is determined based on the experimental conditions: It is determined whether there is a correlation between the performance indicators of targets generated within the same field of view. If there is a correlation, the tree-shaped fully connected network is set to three layers: the first fully connected layer is a single fully connected layer, the number of fully connected layers in the second fully connected network is determined based on the number of fields of view in the experimental conditions, and each upper fully connected layer in the third fully connected network connects to K fully connected layers, where the value of K is determined based on the number of experiments in each field of view in the experimental conditions. If there is no correlation between targets, the tree-shaped fully connected network is set to two layers: the first fully connected layer is a single fully connected layer, and the number of fully connected layers in the second fully connected network is determined based on the number of fields of view and the number of experiments in the experimental conditions. (5) Input the lens parameters of the optical lens into the trained optical lens performance prediction model to obtain the regression value of the lens performance index; (6) Calculate the derivative of the regression value of each lens performance index with respect to the lens parameters to form the changing trend of the lens parameters with respect to the lens performance index.
2. The method for extracting the trend of optical lens performance changes based on gradient analysis according to claim 1, characterized in that, Step (1) specifically includes: setting a preset yield rate, defining the physical range of each lens parameter, uniformly sampling lens parameters to generate simulation samples under the constraint of the physical range of lens parameters, calculating the yield rate based on the simulation results, and determining whether the calculated yield rate is less than the preset yield rate. If the calculated yield rate is less than the preset yield rate, the physical range of lens parameters is reduced; otherwise, the physical range of lens parameters is increased. This process samples lens parameters and their simulation results to generate a simulation dataset.
3. The method for extracting the trend of optical lens performance changes based on gradient analysis according to claim 1, characterized in that, Step (2) specifically includes: deleting values in the sampled values of the modulation transfer function that are irrelevant to the lens performance, so as to obtain lens performance-related values as the task objective.
4. The method for extracting the trend of optical lens performance changes based on gradient analysis according to claim 1, characterized in that, In step (4), the step of iteratively training the optical lens performance prediction model using the training set and adjusting the parameters of the optical lens performance prediction model according to the loss function to obtain a trained optical lens performance prediction model specifically includes: The lens parameters in the training set are input into the optical lens performance prediction model to obtain the lens performance index regression value. The loss function is calculated based on the lens performance index regression value and the task objective in the training set. The parameters of the optical lens performance prediction model are then updated in reverse using the loss function and an optimizer to obtain the trained optical lens performance prediction model.
5. The method for extracting the trend of optical lens performance changes based on gradient analysis according to claim 4, characterized in that, The feature extractor includes a first feature extraction module, a second feature extraction module, and a third feature extraction module connected in sequence. Each of the first, second, and third feature extraction modules includes a convolutional layer, an activation function layer, and a max pooling layer. The convolutional layer, activation function layer, and max pooling layer are connected in sequence. The convolutional layer of the first feature extraction module is connected to the convolutional layer of the second feature extraction module via a bypass connection. The max pooling layer of the first feature extraction module is connected to the max pooling layer of the second feature extraction module via a bypass connection. The convolutional layer of the second feature extraction module is connected to the convolutional layer of the third feature extraction module via a bypass connection. The max pooling layer of the second feature extraction module is connected to the max pooling layer of the third feature extraction module via a bypass connection.
6. The method for extracting performance change trend of optical lens based on gradient analysis according to claim 4, characterized in that, The specific steps for obtaining a trained optical lens performance prediction model include: inputting lens parameters from the training set into the optical lens performance prediction model to obtain lens performance index regression values; calculating a loss function based on the lens performance index regression values and the task objective in the training set; using an optimizer to calculate the gradient of the loss function for each batch of samples and smoothing the gradient; then backpropagating to update the parameters of the optical lens performance prediction model; repeating the training until the loss function is less than a preset loss threshold to obtain a trained optical lens performance prediction model.
7. The method of claim 4 or 6, wherein the method further comprises: The optimizer is the Adam optimizer.
8. The method for extracting the trend of optical lens performance changes based on gradient analysis according to claim 1, 4, or 6, characterized in that, The loss function uses mean squared error, and its expression is: in, It is the first Regression values of lens performance indicators for each sample. It is the first The task objective for each sample It represents the total number of samples in the training set.
9. The method for extracting performance change trend of optical lens based on gradient analysis according to claim 1, wherein, The formula for calculating the derivative is: in, It is the first Regression values of individual lens performance indicators It is the first Lens parameters, The activation values of each hidden layer are denoted as follows: arrive The derivative to be found for right The partial derivatives, when expanded, are Differentiate the activation values of the hidden layers layer by layer.