Complex wavefront detection methods based on phase difference method and deep neural network

Abstract

This invention discloses a complex wavefront detection method based on the phase difference method and deep neural networks, comprising the following steps: collecting image information, simulating wavefront phase difference, training a neural network computational model, and model prediction. The beneficial effects of this invention are that this complex wavefront detection method based on the phase difference method and deep neural networks replaces the nonlinear optimization calculation in existing PD methods with forward parallel computation of the neural network model. That is, a simple forward (one-way) computation mode replaces the original complex iterative computation mode, resulting in a universal wavefront detection algorithm model. The one-way computation mode of the neural network model reduces the solution time of nonlinear optimization iterative computation from minutes to milliseconds, which can better meet the needs of quasi-real-time detection of complex dynamic wavefronts and greatly expands the applicable field of the PD method.

Description

Technical Field

[0001] This invention relates to the field of precise wavefront phase detection technology, and in particular to a complex wavefront detection method based on the phase difference method and deep neural networks. Background Technology

[0002] Accurate wavefront phase detection is a key technology for adaptive optics and high-resolution imaging. In 1979, Gonzalves proposed the Phase Diversity (PD) method. This method introduces an imaging channel with known aberrations (such as defocus) outside the focal plane image, using two (or more) images to eliminate the conjugate uncertainty in phase retrieval, thereby reconstructing the wavefront phase information. The PD method has advantages such as simple optical structure, low cost, and applicability to point targets and extended targets, coherent light and incoherent light applications, and has been widely used in high-resolution optical imaging.

[0003] Currently, the core algorithm used in phase difference wavefront detection technology is the optimization solution for a nonlinear cost function, and its main problem is:

[0004] 1. It is necessary to construct the regularization term by mastering the prior information of the wavefront phase in order to avoid getting trapped in local optima;

[0005] 2. The iterative calculations for nonlinear optimization solutions are extremely time-consuming.

[0006] Therefore, the currently used PD method cannot adequately meet the needs of real-time detection of dynamically changing wavefronts. In view of this, in-depth research was conducted on the above-mentioned problems, which led to this case. Summary of the Invention

[0007] The purpose of this invention is to solve the above-mentioned problems by designing a complex wavefront detection method based on phase difference method and deep neural network, which solves the existing background technical problems.

[0008] The technical solution of the present invention to achieve the above objectives is as follows: a complex wavefront detection method based on phase difference method and deep neural network, comprising the following steps: collecting image information, simulating wavefront phase difference, training neural network calculation model, and model prediction;

[0009] Image information collection: Collect focal plane and out-of-focus image information of the same imaging target;

[0010] Wavefront phase difference simulation: Based on the Kolmogorov atmospheric turbulence model, wavefront phase difference samples are simulated. The point spread function (PSF) of the focal plane and the defocus plane is calculated using the wavefront phase difference samples. The focal plane and the defocus image of the corresponding wavefront phase difference sample are generated from the PSF. The focal plane and the defocus image and the corresponding wavefront phase difference constitute the training sample pair of the neural network.

[0011] Training a neural network computational model involves first transforming the data, and then inputting the results of the data transformation into the hybrid neural network for training.

[0012] Data Transformation: Fourier transforms are performed on the focal plane image and the out-of-focus image to obtain their frequency domain representations. and ;pass and Calculated , and ;

[0013] Hybrid network training: real part and the virtual part ,as well as , As input, the Zernikal polynomial fitting result of the wavefront phase difference is used as output to train the hybrid neural network;

[0014] Model prediction: Collect focal plane and defocus images, input them into a hybrid neural network through data transformation, and obtain the Zernike polynomial fitting result of the wavefront phase difference as the prediction result.

[0015] The point spread function (PSF) of the focal plane and defocus plane simulated by the wavefront phase difference is calculated using the phase difference method imaging optical path.

[0016] The number of samples used for wavefront phase difference simulation training is no less than 20,000.

[0017] When performing data transformation during the training of the neural network computation model , and The calculation method is as follows ,

[0018] In the calculation formula for the data transformation, '*' represents taking the complex conjugate, and '||' represents the modulo operation.

[0019] The complex wavefront detection method based on phase difference and deep neural networks, developed using the technical solution of this invention, trains a neural network computational model with simulated wavefront phase samples, enabling it to obtain a universal mapping relationship between the focal plane image and the defocus image to calculate wavefront phase information. This computational model then provides real-time approximation of the dynamic complex wavefront phase. Its main advantages are as follows:

[0020] 1. The forward parallel computation of the neural network model replaces the nonlinear optimization computation in the existing PD method. That is, a simple forward (one-way) computation mode replaces the original complex iterative computation mode, resulting in a universal wavefront detection algorithm model.

[0021] 2. The unidirectional computation mode of the neural network model reduces the solution time of nonlinear optimization iteration from minutes to milliseconds, which can better meet the needs of quasi-real-time detection of complex dynamic wavefronts and greatly expand the application field of the PD method. Attached Figure Description

[0022] Figure 1 This is a phase difference (PD) imaging optical path diagram of the complex wavefront detection method based on phase difference method and deep neural network described in this invention.

[0023] Figure 2 This is a flowchart of the deep neural network wavefront detection process based on PD, which is part of the complex wavefront detection method based on phase difference and deep neural networks described in this invention. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings:

[0025] This invention discloses a novel method for predicting complex wavefront phase aberrations (such as the phase of atmospheric turbulence on Earth) using deep neural networks by utilizing focal plane and out-of-focus image information of the same imaging target.

[0026] This method trains a neural network computational model using simulated wavefront phase samples, enabling it to obtain a universal mapping relationship between the focal plane image and the defocus image to calculate wavefront phase information. The computational model then approximates dynamically complex wavefront phases in real time. The method mainly consists of the following three steps:

[0027] I. Wavefront Phase Aberration Simulation

[0028] 1.1 Generate simulated wavefront phase difference samples (generally no less than 20,000 samples) based on the Kolmogorov atmospheric turbulence model.

[0029] 1.2 Press Figure 1 The principle is to calculate the point spread function (PSF) of the focal plane and the defocus plane using wavefront phase difference samples.

[0030] 1.3 Generate the focal plane and out-of-focus image of the corresponding wavefront phase difference sample from the above PSF.

[0031] 1.4 The focal plane and the out-of-focus image, along with their corresponding wavefront phase differences, constitute the training sample pairs for the neural network.

[0032] II. Model Training

[0033] 2.1 Data Transformation:

[0034] The frequency domain representations of the in-focus and out-of-focus images are obtained by performing Fourier transforms on them: and ;Depend on and The following results were obtained. , and : Where '*' represents the complex conjugate and '||' represents the modulo operation;

[0035] 2.2 with real part and the virtual part ,as well as , As input, the Zernikal polynomial fitting result of the wavefront phase difference is used as output for training. Figure 2 Hybrid neural networks in [the context of] ...

[0036] III. Model Prediction

[0037] use Figure 1 The optical system shown acquires focal plane and defocus images, and after data transformation in section 2.1, inputs them into a hybrid neural network to obtain the Zernike polynomial fitting result of the wavefront phase difference as an estimate.

[0038] See Table 1 for a comparison of the improved data:

[0039] Table 1. Comparison of key effects before and after improvement

[0040]

[0041] This complex wavefront detection method based on phase difference and deep neural networks trains a neural network computational model using simulated wavefront phase samples, enabling it to obtain a universal mapping relationship between the focal plane image and the defocus image to calculate wavefront phase information. This computational model then approximates the dynamic complex wavefront phase in real time. The forward parallel computation of the neural network model replaces the nonlinear optimization computation in existing PD methods; that is, a simple forward (one-way) computation mode replaces the original complex iterative computation mode, resulting in a universal wavefront detection algorithm model. The one-way computation mode of the neural network model reduces the solution time of nonlinear optimization iteration computation from minutes to milliseconds, which can better meet the needs of near real-time detection of complex dynamic wavefronts and greatly expands the application field of the PD method.

[0042] The above technical solutions only embody the preferred technical solutions of the present invention. Any modifications that may be made by those skilled in the art to certain parts thereof embody the principles of the present invention and fall within the protection scope of the present invention.

Claims

1. A complex wavefront detection method based on the phase diversity method and a deep neural network, characterized in that, The method comprises the following steps: collecting image information, simulating wavefront phase difference, training a neural network calculation model, and model prediction. The image information is collected, including the focal plane and defocus image information of the same imaging target. The wavefront phase difference samples are simulated based on the Kolmogorov atmospheric turbulence model, the point spread function (PSF) of the focal plane and defocus plane is calculated based on the wavefront phase difference samples, the focal plane and defocus image corresponding to the wavefront phase difference samples is generated based on the PSF, and the focal plane and defocus image and the corresponding wavefront phase difference constitute a training sample pair of the neural network. The training of the neural network calculation model is performed after data transformation. Data transformation: the in-focus and out-of-focus images are Fourier transformed to obtain their frequency domain representations: and ; by and calculating , and ; Mixed network training: take the real part and the imaginary part of the complex number , and , as input, and the Zernike polynomial fitting result of the wavefront phase difference as output, to train the mixed neural network; The focal plane and defocus image is collected, the data is transformed, the hybrid neural network is input, and the Zernike polynomial fitting result of the wavefront phase difference is obtained as the prediction result. The training of the neural network computing model is performed with data transformation, , and The calculation method is , In the calculation formula of the data transformation, '*' represents the complex conjugate, and '||' represents the modulus operation.

2. The complex wavefront sensing method based on phase diversity method and deep neural network according to claim 1, wherein, The point spread function (PSF) of the focal plane and defocus plane of the wavefront phase difference simulation is calculated through the phase difference imaging light path.

3. The complex wavefront sensing method based on phase diversity method and deep neural network according to claim 2, wherein, The number of sample training of the wavefront phase difference simulation is not less than 20,000 samples.

Images