Deep learning optical imaging system design method based on complex function neural network

Abstract

The application provides a kind of deep learning optical imaging system design method based on complex function neural network, comprising the following steps: step one, object distance, system focal length, F number as input, according to the ideal optical system imaging relationship, determine image distance, imaging scale ratio;Step two, according to the imaging scale ratio and the object image of ideal optical system, the image of ideal optical system is calculated;Step three, construct complex function generator network model, step four, construct discriminator network model, step five, after the establishment, training of the above-mentioned complex function neural network deep learning model, step six, use phase.The deep learning optical imaging system design method based on complex function neural network provided by the application solves the correlation between the advantages and disadvantages of the design result and the experience value of the designer, without the designer providing initial structure parameters;Greatly improve the computing ability and optimization speed.

Description

Technical Field

[0001] This invention relates to the field of catadioptric optical imaging technology, and more specifically to a deep learning optical imaging system design method based on complex function neural networks. Background Technology

[0002] Optical system design, or simply optical design, is the process of determining the design specifications that meet the user's requirements, including the performance parameters, dimensions, first-order parameters, and structure of each optical group of the optical system.

[0003] The traditional optical system design process usually includes: (1) design or selection of the initial structure; (2) dozens or even hundreds of Maxwell electromagnetic field simulations are required; (3) based on the design indicators, an optimization function is established to optimize the optical design. This type of optical design process has been solidified into the basic thinking of optical designers and R&D personnel, but it has fallen into a technical bottleneck. The main inherent technical disadvantages of this method are as follows: (1) dozens or even hundreds of Maxwell electromagnetic field simulations are required, which is time-consuming and computationally expensive. The optimization algorithm and partial differential iterative optimization algorithm in lens design consume a lot of CPU threads. Optical imaging design involving interference models and diffraction models not only consumes CPU threads, but also requires real-time memory space of the computer for multi-dimensional matrix operations. (2) Designers must spend a lot of time and energy on preparation, modeling and optimization. Deviations in modeling conditions will cause the design to fail completely, and it is difficult to find the cause of the optical design failure. (3) In the traditional design of optical imaging systems, the optimization algorithms used are based on the least squares method or improved algorithms of the least squares method. During the optimization process, it is easy to get stuck in local solutions and it is difficult to find the global optimal solution.

[0004] Deep learning-based artificial intelligence technology not only provides new ideas in algorithms but also offers new means in terms of computing hardware such as graphics processing units (GPUs) or tensor processing units (TPUs). It has significantly improved both the algorithm (software) and parallel computing (hardware) aspects of many optical algorithms. Since 2012, it has been widely applied in computer vision and image processing, primarily through a data-driven approach that reduces the reliance on known boundary conditions. Currently, it is divided into two main categories: data-driven deep learning and deep learning based on complex function neural networks. This data-driven approach quickly encountered bottlenecks in various optical technologies, revealing numerous problems such as incorrect training, overtraining, and limited applicability. Deep learning algorithms based on complex function neural networks, built upon actual physical models, have broader application value. In particular, compared to traditional data-driven deep learning, algorithms based on complex function neural networks offer faster optimization speeds and more accurate optimization results.

[0005] However, in recent years, deep learning models have begun to be used to guide the design, construction, and training of networks for deep optics, and are currently mainly applied to the design optimization of devices and systems such as photonic computing. This patent proposes for the first time a design optimization method for lens-type catadioptric optical element imaging systems based on complex function neural networks. Drawing on breakthroughs in deep learning algorithms based on complex function neural networks, the design methods for optical imaging systems based on deep learning have a clearer direction for evolution. Applying deep learning algorithms based on complex function neural networks to the design optimization of lens-type catadioptric imaging systems fundamentally improves upon the shortcomings of least squares optimization algorithms in traditional optical imaging systems, significantly increasing optimization speed and accuracy.

[0006] Therefore, how to solve the design optimization problem of traditional optical imaging systems based on the least squares method, and realize a faster and more convenient design process for optical imaging systems, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide a design method for a deep learning optical imaging system based on complex function neural networks, addressing the problems in the prior art.

[0008] Therefore, the above-mentioned objectives of the present invention are achieved through the following technical solutions:

[0009] The design method for deep learning optical imaging systems based on complex function neural networks includes the following steps:

[0010] Step 1: Using the object distance, system focal length, and F-number as inputs, determine the image distance and imaging scaling ratio based on the imaging relationship of an ideal optical system.

[0011] Step 2: Calculate the image-side image of the ideal optical system based on the imaging scaling ratio and the object-side image of the ideal optical system.

[0012] Step 3: Construct a complex function generator network model, which includes a first layer network, a second layer network, ..., an Nth layer network. The first layer network consists of the curvature, refractive index, and physical spacing of the first lens surface. The propagation to the second layer network consists of the curvature, refractive index, and physical spacing of the second lens surface. Finally, the propagation to the Nth layer network consists of the curvature, refractive index, and physical spacing of the Nth lens surface.

[0013] Step 4: Construct the discriminator network model. The image-side image output by the complex function generator network model is input into the discriminator network model, which judges the similarity between the image-side image and the image-side image of the ideal optical system. The judgment threshold of the discriminator network is determined by the design parameters. When the discriminator network model judges that the similarity between the image-side image and the image-side image of the ideal optical system reaches the threshold, it outputs "true", stops feeding back the recovery function generator network model, and ends the iteration. When the discriminator network model judges that the similarity between the image-side image and the image-side image of the ideal optical system does not reach the threshold, it outputs "false" and feeds this signal back to the complex function generator network model to continue iterative optimization of the complex function generator network parameters in Step 3. The object-side image of the ideal optical system is input into the discriminator network model, which judges whether further optimization calculation is needed, and finally outputs the specific values ​​of the curvature, refractive index, and physical spacing of lens surfaces 1 to N.

[0014] Step 5: After establishing and training the deep learning model of the complex function neural network, the first-order parameters object distance, system focal length, and F-number are discretized and sampled. The discretized object distance, system focal length, and F-number are sequentially combined as a set of inputs. Steps 1 to 4 are repeated until all discretized object distances, system focal lengths, and F-numbers are combined. Specifically, the object distance is sampled at intervals of 0.1m to 1m, at intervals of 1m to 10m, at intervals of 10m to 100m, and at intervals of 10m to 100m. Sampling is performed at intervals of 100m for ~3000m, at intervals of 10mm for system focal length from 10mm to 50mm, at intervals of 50mm for system focal length from 50mm to 500mm, and at intervals of 0.5 for F-number from 1.5 to 10. Then, a series of discretized first-order parameters are input into a deep learning algorithm of a complex function neural network. The curvature, refractive index, and physical spacing of lens surfaces 1 to N are stored in a database as the initial structure of the optical imaging system in the usage stage.

[0015] Step Six, the usage phase: In the complex function neural network deep learning algorithm that has been trained and verified for accuracy, a set of optical imaging system design specifications and first-order parameters are input, including object distance, system focal length, and F-number. First, similar structures are searched in the database established in Step Five. That is, the structure with the smallest difference between the three values ​​of object distance, system focal length, and F-number and the three values ​​of the input first-order parameters is identified as a similar structure and used as the initial structure. After design and optimization by the complex function neural network deep learning algorithm, the specific values ​​of the curvature of lens surfaces 1 to N, the refractive index of lens surfaces 1 to N, and the physical distance between lens surfaces 1 to N are output, thus completing the design of the optical imaging system.

[0016] While adopting the above technical solutions, the present invention may also adopt or combine the following technical solutions:

[0017] As a preferred technical solution of the present invention: the principle of constructing the complex function generator network in step three is as follows: the propagation of light waves in a medium is specifically manifested as a phase change. In the same medium, light waves follow diffraction propagation. During the propagation from lens surface n to lens surface n+1, the refractive index is... The physical spacing is The propagation formula is as shown in equation (1).

[0018] (1)

[0019] In the formula, Represents wavelength, ( ) represents three-dimensional position coordinates, R ;

[0020] When the propagation reaches the lens surface n+1, the refractive indices of the two media are respectively and When the refractive index changes, the complex amplitude distributions of the light wave on the plane immediately in front of the lens surface and on the plane immediately behind the lens surface are respectively...

[0021] (2)

[0022] In the formula, A represents the amplitude of the light wave, and (x, y) represents the two-dimensional coordinates. It is a constant value, related to the physical distance D. It is phase.

[0023] Therefore, the phase change of the incident light wave by the lens is as follows:

[0024] (3)

[0025] Light waves in The phase change produced by the lens surface at that point is

[0026] (4)

[0027] In the formula, and These are the distances between the lens surface and the front and rear planes, respectively, and they vary with... It changes with the coordinates. It is the refractive index of the medium in front of the lens surface. It is the refractive index of the medium behind the lens. Represents wavelength, and The value is determined according to the lens surface shape formula.

[0028] Thus, the model of light wave propagating from lens surface n to lens surface n+1 is established. By passing this model sequentially, the entire complex function generator network model can be established.

[0029] As a preferred technical solution of the present invention: the complex function generator network model in step three is trained, optimized, and verified for accuracy. When training the complex function generator network model, 80% of the data in the neural network training dataset is used to replace the object-side image of the ideal optical system as the input of the complex function generator network model to train and optimize the complex function generator network model and the discriminator network model. When the training is completed, the remaining 20% ​​of the neural network training dataset is used as verification data to verify the accuracy of the trained network and obtain effective model parameters for optical imaging system design optimization.

[0030] As a preferred technical solution of the present invention: In step four, the discriminator network model includes an input layer, a network layer and an output layer, which are arranged in sequence. The network layer of the discriminator network model includes a convolutional layer, a compression layer, a nonlinear activation layer, a one-dimensional layer and a fully connected layer.

[0031] As a preferred technical solution of the present invention: In step six, when a set of first-order parameters, including object distance, system focal length, and F-number, are input, a set of similar structures are taken from the database established in step five as the initial structure of the optical imaging system for optimization design. The similar structure refers to the initial structure corresponding to the set in step five where the first-order parameters object distance, system focal length, and F-number have been discretized and sampled. It is only necessary to find the set of three parameters that have the smallest difference with the three parameters that have been sampled.

[0032] Compared with existing technologies, the deep learning-based optical imaging system design method based on complex function neural networks provided by this invention utilizes deep learning neural network models with practical optical physics significance. Algorithmically, it does not rely on statistical calculation methods from computer science, greatly improving the matching degree between deep learning networks and practical engineering problems in optical imaging system design. It resolves the correlation between the quality of design results and the designer's experience, eliminating the need for the designer to provide initial structural parameters. In terms of hardware, it fully utilizes the computing hardware of graphics processing units (GPUs) or tensor processing units (TPUs), significantly improving computing power and optimization speed. This method greatly enhances the matching degree between deep learning networks and practical engineering problems in optical imaging system design, possessing advantages such as universality, convenience, and a more user-friendly experience. It significantly simplifies the design difficulty of optical imaging systems, making the design process faster, more convenient, and more user-friendly, thus having pioneering significance in the technological development of optical imaging system design. Attached Figure Description

[0033] Figure 1This is a flowchart of the deep learning optical imaging system design method based on complex function neural networks according to the present invention;

[0034] Figure 2 This is the object-side image (O) of the ideal optical system of the present invention;

[0035] Figure 3 The image-side image of the ideal optical system of the present invention ( );

[0036] Figure 4 This is a partial data diagram of the neural network training dataset (DATA) used in this invention;

[0037] Figure 5 This is a schematic diagram of the optical imaging system designed for this invention. Detailed Implementation

[0038] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0039] The present invention provides a deep learning optical imaging system design method based on complex function neural networks, including object distance l, system focal length f, F-number F#, and image distance. Imaging scaling ratio R, object-side image O of the ideal optical system, and image-side image of the ideal optical system. Complex function generator network model GN, curvature of lens surface 1 The refractive index of lens surface 1 Physical spacing of lens surface 1 The curvature of lens surface N The refractive index of lens surface N Physical spacing of lens surface N Image I, discriminator network model DN, neural network training dataset DATA, where:

[0040] The object distance l, the system focal length f, and the F-number F# are input into a deep learning algorithm based on a complex function neural network. Using ideal imaging relationships, the image distance is determined. The imaging scaling ratio R is used to determine the object-side image O and the image-side image of the ideal optical system. Simultaneously, the object-side image O of the ideal optical system is input into the complex function generator network model GN. After passing through the first layer of the network, it is generated by the parameter curvature of lens surface 1. Refractive index and physical distance It is constructed; then propagated to the second layer network, by the curvature of the lens surface 2. Refractive index and physical distance It is constructed; finally, it propagates to the Nth layer network, determined by the curvature parameters of the lens surface N. Refractive index and physical distance Composition; output image-side image I. Image-side image I and the image-side image of the ideal optical system. The input is fed into the discriminator network model DN, which determines whether further optimization is needed. The final output is the curvature of each lens surface from 1 to N. Refractive index of lens surfaces 1~N Physical spacing between lens surfaces 1 to N The specific values ​​are as follows. During the training phase, the neural network training dataset DATA replaces the object-side image O of the ideal optical system and is input into the complex function generator network model GN. The complex function generator network model GN and the discriminator network model DN are trained to obtain effective model parameters for optical imaging system design optimization. The neural network training dataset DATA is a series of images collected from the internet, specifically as follows: Figure 4 .

[0041] A deep learning-based optical imaging system design method based on complex function neural networks is presented, and its steps are as follows:

[0042] Step 1: Calculate the first-order parameters of the optical imaging system. Given the first-order parameters object distance l, system focal length f, and F-number F#, calculate the image distance using the object distance l, system focal length f, and F-number F# according to the imaging relationship of an ideal optical imaging system. Image scaling ratio R.

[0043] Step two: Based on the imaging scaling ratio R and the object-side image O of the ideal optical system, the image-side image of the ideal optical system can be calculated. .

[0044] Step 3: Construct the complex function generator network model GN. The complex function generator network model GN takes an object-space image O of an ideal optical system and design specifications as input, and outputs an image-space image I starting from random data.

[0045] Step four: Construct the discriminator network model DN. The image-space image I output by the complex function generator network model GN is input into the discriminator network model DN, which then determines whether the image-space image I matches the image-space image of the ideal optical system. The similarity. When the discriminator network model DN determines the image-side image I and the image-side image of the ideal optical system. When the similarity reaches the threshold, output "true" and stop iteratively optimizing the complex function generator network model GN. When the discriminator network model DN determines that the image-side image I is similar to the image-side image of the ideal optical system... If the similarity does not reach the threshold, "false" is output, and this signal is fed back to the complex function generator network model GN to continue iterative optimization of the complex function generator network model GN in step three.

[0046] The construction of the complex function generator network model requires a training phase. 80% of the data in the neural network training dataset DATA is used as input to the complex function generator network model GN, training and optimizing both GN and the discriminator network model DN. Once training is complete, the remaining 20% ​​of the neural network training dataset DATA is used as validation data to verify the accuracy of the trained network.

[0047] Step 5: After establishing and training the deep learning model of the complex function neural network, the first-order parameters object distance l, system focal length f, and F number F# are discretized and sampled. The discretized object distance l, system focal length f, and F number F# are arranged and combined in sequence as a set of inputs. Steps 1 to 4 are repeated until all discretized object distance l, system focal length f, and F number F# are arranged and combined. Specifically, the object distance l is sampled at intervals of 0.1m to 1m, 1m to 10m, 10m, 10m, and 3000m. The system focal length f is sampled at intervals of 10mm to 50mm, 50mm, and 10mm. The F number F# is sampled at intervals of 0.5 from 1.5 to 10. Then, a series of discretized first-order parameters are input into the deep learning algorithm of the complex function neural network. The curvature r of the corresponding lens surfaces 1 to N, the refractive index n of the lens surfaces 1 to N, and the physical distance D of the lens surfaces 1 to N are stored in a database as the initial structure of the optical imaging system in the usage stage. This can greatly improve the optimization speed and eliminate the need for the designer to input the initial structure of the optical imaging system, making it more convenient to use.

[0048] Step Six, the usage phase: In the complex function neural network deep learning algorithm that has been trained and verified for accuracy, a set of optical imaging system design specifications and first-order parameters are input, including object distance, system focal length, and F-number. First, similar structures are searched in the database established in Step Five. That is, the structure with the smallest difference between the three values ​​of object distance, system focal length, and F-number and the three values ​​of the input first-order parameters is identified as a similar structure and used as the initial structure. After optimization by the complex function neural network deep learning algorithm, the specific values ​​of curvature r, refractive index n, and physical distance D of each lens surface 1 to N are output, thus completing the design of the optical imaging system.

[0049] Unlike conventional data-driven deep learning optical design methods, the present invention provides a deep learning optical imaging system design method based on complex function neural networks, which has significant advantages over existing technologies.

[0050] (1) Design optimization algorithms based on deep learning fully utilize the computing power of computers to improve the computing speed of optimization design;

[0051] (2) It solves the problem of the correlation between the quality of the design results and the designer's direct experience value. The designer does not need to master the complex aberration theory and aberration optimization theory to design optimization functions. At the same time, the designer does not need to provide initial structural parameters, making the design process simpler.

[0052] (3) It solves the limitations of the optimization algorithm based on the least squares method in the traditional optical imaging system design process, and the design results are more accurate.

[0053] Therefore, compared with the prior art, the deep learning optical imaging system design method based on complex function neural network of the present invention has the advantages of universality, convenience and more user-friendly experience, which greatly simplifies the design difficulty of optical imaging system, thereby making the design process of optical imaging system faster, more convenient and more user-friendly.

[0054] Example 1

[0055] like Figures 1-5 As shown, the present invention provides a deep learning optical imaging system design method based on complex function neural networks.

[0056] The optical imaging system is designed with the following specifications: wavelength 530nm, aperture 40mm, object height 20mm, object distance 500mm, system focal length 170mm, and F-number 8.5.

[0057] A deep learning-based optical imaging system design method based on complex function neural networks is presented, and its steps are as follows:

[0058] Step 1: Combining Figure 1 Calculate the first-order parameters of the optical imaging system. The first-order parameter, object distance, is known. The system focal length f and F-number F# are determined based on the imaging relationship of an ideal optical imaging system, using the object distance. The image distance can be calculated from the system focal length f and the F-number F#. Image scaling ratio R.

[0059] Step Two, Combining Figure 2 and Figure 3 Based on the imaging scaling ratio R and the object-space image O of the ideal optical system, such as Figure 2 The image-side image of an ideal optical system can be calculated. like Figure 3 .

[0060] Step 3: Construct the complex function generator network model GN. The complex function generator network model GN takes as input an object-space image O of an ideal optical system and design specifications of wavelength 530nm, aperture 40mm, and object height 20mm. Starting from random data, the complex function generator network model GN outputs an image-space image. .

[0061] Step 4: Construct the discriminator network model DN. The image output by the complex function generator network model GN is shown. The image is input into the discriminator network model DN, which then determines the image's orientation. Image of an ideal optical system The similarity. When the discriminator network model DN judges the similarity of the image. Image of an ideal optical system When the similarity reaches the threshold, the output is "true," and the iterative optimization of the complex function generator network model GN stops. When the discriminator network model DN judges the image as... Image side of an ideal optical system If the similarity does not reach the threshold, "false" is output, and this signal is fed back to the complex function generator network model GN to continue iterative optimization of the complex function generator network model GN in step three.

[0062] In step four, combined with Figure 4 During the training phase, 80% of the data in the neural network training dataset DATA is used as input to the complex function generator network model GN to train and optimize both the complex function generator network model GN and the discriminator network model DN. Once training is complete, the neural network training dataset DATA is then used as follows: Figure 4 The remaining 20% ​​is used as validation data to verify the accuracy of the trained network.

[0063] Step 5: After verifying the deep learning model of the complex function neural network in Step 5, the first-order parameters object distance l, system focal length f, and F number F# are discretized and sampled. The discretized object distance l, system focal length f, and F number F# are arranged and combined in sequence as a set of inputs. Steps 1 to 4 are repeated until all discretized object distance l, system focal length f, and F number F# are arranged and combined. Specifically, the object distance is sampled at intervals of 0.1m to 1m, 1m to 10m, 10m, 10m, and 3000m. The system focal length is sampled at intervals of 10mm to 50mm, 50mm, and 1.5mm. The F number is sampled at intervals of 0.5. Then, a series of discretized first-order parameters are input into a deep learning algorithm of a complex function neural network. The curvature, refractive index, and physical spacing of lens surfaces 1 to N are stored in a database as the initial structure of the optical imaging system in the usage stage, which can greatly improve the optimization speed.

[0064] Step Six: Usage Phase. In the complex function neural network deep learning algorithm, which has been trained and verified for accuracy, a set of first-order parameters for the optical imaging system is input: object distance 500mm, system focal length 170mm, F-number 8.5. An initial structure corresponding to an object distance of 500mm, system focal length 150mm, and F-number 8.5 is selected from the database as the initial structure before optimization. After optimization by the complex function neural network deep learning algorithm, the specific values ​​of curvature r, refractive index n, and physical distance D for each of lens surfaces 1-6 are output, completing the design of the optical imaging system. The specific parameters are as follows: =100mm, =1.755, =5.4mm; =-1247.6mm, =1, =5.0mm; =-122.5mm, =1.643, =2.9mm; =113.1mm, =1, =6.4mm; =331.4mm, =1.655, =6.5mm; =-102.3mm, =1, =246.2mm, consisting of 3 lenses composed of 6 lens surfaces, the schematic diagram of the optical imaging system is as follows. Figure 5 .

[0065] The above specific embodiments are used to explain and illustrate the present invention, and are only preferred embodiments of the present invention, not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A design method for a deep learning optical imaging system based on complex function neural networks, comprising the following steps: Step 1: Using the object distance, system focal length, and F-number as inputs, determine the image distance and imaging scaling ratio based on the imaging relationship of an ideal optical system. Step 2: Calculate the image-side image of the ideal optical system based on the imaging scaling ratio and the object-side image of the ideal optical system. Step 3: Construct a complex function generator network model, which includes a first layer network, a second layer network, ..., an Nth layer network. The first layer network consists of the parameters curvature, refractive index, and physical spacing of the first lens surface. The propagation extends to the second layer network, which consists of the parameters curvature, refractive index, and physical spacing of the second lens surface. Finally, the propagation extends to the Nth layer network, which consists of the parameters curvature, refractive index, and physical spacing of the Nth lens surface. Step 4: Construct the discriminator network model. The image-side image output by the complex function generator network model is input into the discriminator network model, which judges the similarity between this image-side image and the image-side image of the ideal optical system. The judgment threshold of the discriminator network is determined by the design parameters. When the discriminator network model determines that the similarity between the image-side image and the image-side image of the ideal optical system reaches the threshold, it outputs "true," stops feeding back the recovery function generator network model, and ends the iteration. When the discriminator network model determines that the similarity between the image-side image and the image-side image of the ideal optical system does not reach the threshold, it outputs "false," and feeds back this "false" signal to the complex function generator network. The model continues to iteratively optimize the parameters of the complex function generator network from step three. During the training of the complex function generator network model, 80% of the data in the neural network training dataset is used to replace the object-side image of the ideal optical system as the input to the complex function generator network model, thereby training and optimizing the complex function generator network model and the discriminator network model. When training is complete, the remaining 20% ​​of the neural network training dataset is used as validation data to verify the accuracy of the trained network, obtain effective model parameters, and optimize the optical imaging system design. The final output is the specific values ​​of the curvature, refractive index, and physical spacing of lens surfaces 1 to N. Step 5: After establishing and training the deep learning model of the complex function neural network, the first-order parameters object distance, system focal length, and F-number are discretized and sampled. The discretized object distance, system focal length, and F-number are sequentially combined as a set of inputs. Steps 1 to 4 are repeated until all discretized object distances, system focal lengths, and F-numbers are combined. Specifically, the object distance is sampled at intervals of 0.1m to 1m, at intervals of 1m to 10m, at intervals of 10m to 100m, and at intervals of 10m to 100m. Sampling is performed at intervals of 100m for ~3000m, at intervals of 10mm for system focal length from 10mm to 50mm, at intervals of 50mm for system focal length from 50mm to 500mm, and at intervals of 0.5 for F-number from 1.5 to 10. Then, a series of discretized first-order parameters are input into a deep learning algorithm of a complex function neural network. The curvature, refractive index, and physical spacing of lens surfaces 1 to N are stored in a database as the initial structure of the optical imaging system in the usage stage. Step six, the usage phase: In the complex function neural network deep learning algorithm that has been trained and verified for accuracy, a set of optical imaging system design indicators and first-order parameters, including object distance, system focal length, and F-number, are input. First, similar structures are searched in the database established in step five and used as initial structures. After design and optimization by the complex function neural network deep learning algorithm, the specific values ​​of the curvature of lens surfaces 1 to N, the refractive index of lens surfaces 1 to N, and the physical spacing of lens surfaces 1 to N are output, thus completing the design of the optical imaging system.

2. The design method for a deep learning optical imaging system based on a complex function neural network as described in claim 1, characterized in that: The principle of constructing the complex function generator network in step three is as follows: the propagation of light waves in a medium is specifically manifested as a phase change. In the same medium, light waves follow diffraction propagation. During the propagation from lens surface n to lens surface n+1, the refractive index changes. The physical spacing is The propagation formula is as shown in equation (1). (1) In the formula, Represents wavelength, ( ) represents three-dimensional position coordinates, R , ; When the propagation reaches the lens surface n+1, the refractive indices of the two media are respectively and When the refractive index changes, the complex amplitude distributions of the light wave on the plane immediately in front of the lens surface and on the plane immediately behind the lens surface are respectively... (2) In the formula, A represents the amplitude of the light wave, and (x,y) represents the two-dimensional coordinates. It is a constant value, related to the physical distance D. It is phase; Therefore, the phase change of the incident light wave by the lens is as follows: (3) Light waves in The phase change produced by the lens surface at that point is (4) In the formula, and These are the distances between the lens surface and the front and rear planes, respectively, and they vary with... It changes with the coordinates. It is the refractive index of the medium in front of the lens surface. It is the refractive index of the medium behind the lens. Represents wavelength, and The value is determined according to the lens surface shape formula; Thus, the model of light wave propagating from lens surface n to lens surface n+1 is established. By passing this model sequentially, the entire complex function generator network model can be established.

3. The design method for a deep learning optical imaging system based on a complex function neural network as described in claim 1, characterized in that: In step four, the discriminator network model includes an input layer, a network layer, and an output layer, arranged in sequence. The network layer of the discriminator network model includes a convolutional layer, a compression layer, a nonlinear activation layer, a one-dimensional layer, and a fully connected layer.

4. The design method for a deep learning optical imaging system based on a complex function neural network as described in claim 1, characterized in that: In step six, when a set of first-order parameters, including object distance, system focal length, and F-number, are input, a set of similar structures is taken from the database established in step five as the initial structure of the optical imaging system for optimization design. The similar structure refers to the initial structure corresponding to the set of three parameters that have been discretized and sampled in step five. The three parameters are the required three parameters and the three sampled parameters with the smallest difference.

Images