Intelligent Wavefront Detection Parameter Migration Method and System Based on Equivalent Fresnel Number

Abstract

The present invention discloses an intelligent wavefront detection parameter migration method and system based on the equivalent Fresnel number, belonging to the field of laser wavefront detection. Aiming at the problem of network model migration of intelligent wavefront correction methods under different system parameters, by training a wavefront restoration deep learning network model and using this model for laser system parameter migration, there is no need to retrain for a specific laser system. It is possible to combine a large number of simulation samples and a small number of experimental samples to solve the problem of difficult acquisition of actual training samples and improve the adaptability and robustness of the wavefront restoration deep learning network model.

Description

Technical Field

[0001] The present invention belongs to the field of laser wavefront detection, and particularly relates to an intelligent wavefront detection parameter migration method and system based on an equivalent Fresnel number. Background Art

[0002] With the wide application of deep learning, intelligent wavefront detection technology based on deep learning has gradually been used in laser wavefront detection. The most extensive one is to collect the light spots at two positions of the focus and defocus, as Figure 1 shown, so as to invert the distorted wavefront of the laser field by means of a deep learning network (i.e., a wavefront restoration deep learning network model). For example, Ma Huimin et al. from Anhui Agricultural University proposed an adaptive optical correction method based on a convolutional neural network, and YUNCHENG JIN et al. from Zhejiang University proposed a wavefront reconstruction method for biological observation, etc. The existing intelligent wavefront detection or correction systems are based on supervised or unsupervised deep learning networks, and establish a mapping relationship between the wavefront and the detector light spot through the network. However, the network is highly targeted and can only match the scenarios under a single system parameter. When the laser system parameters change, the network needs to be retrained and constructed again. The construction of the network not only has a high time cost, but also is time-consuming and laborious in data generation and acquisition, which is not conducive to engineering applications. Summary of the Invention

[0003] The object of the present invention is to address the problem of network model migration of the intelligent wavefront correction method under different system parameters, and propose an intelligent wavefront detection parameter migration method and system based on an equivalent Fresnel number, which can enhance the generalization of the intelligent wavefront correction method. For the training of the wavefront restoration deep learning network model, laser system parameter migration can be performed through calculation, and there is no need to retrain for a specific laser system. A combination of a large number of simulation samples and a small number of experimental samples can be used to solve the problem of difficult acquisition of actual training samples, and improve the adaptability and robustness of the wavefront restoration deep learning network model.

[0004] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0005] An intelligent wavefront detection parameter migration method based on an equivalent Fresnel number, comprising the following steps:

[0006] 1) Collect training data according to the parameters of the known laser system A, and use the training data to train the wavefront restoration deep learning network model;

[0007] 2) According to the parameters of the laser system A, calculate the equivalent Fresnel number Nf of the laser system A, and select the appropriate number of pixels Dim1 of the focus light spot and the appropriate number of pixels Dim2 of the defocus light spot of the laser system A;

[0008] 3) For the laser system B to be measured, by adjusting the defocus position of the laser system B, make the equivalent Fresnel number of the laser system B equal to that of the laser system A, and calculate the appropriate number of pixels Dim1' of the focal spot of the laser system B and the appropriate number of pixels Dim2' of the defocused spot according to the Dim1 and Dim2;

[0009] 4) For the laser system B, adjust the detection position of the defocus detector according to the defocus position calculated in step 3), intercept the pixel sizes of Dim1' and Dim2' with the spot centroid as the center in the focal point and the defocus detector, and perform resampling, and then input the resampled data into the wavefront reconstruction deep learning network model to obtain the wavefront phase of the laser system B.

[0010] Preferably, the training data includes the focal point, the defocused spot, and the distorted wavefront phase distribution.

[0011] Preferably, the wavefront reconstruction deep learning network model is based on CNN or CGAN.

[0012] Preferably, calculate the corresponding equivalent Fresnel number Nf according to the beam aperture, wavelength, focal length, and defocus position of the laser system A.

[0013] Preferably, select the appropriate number of pixels Dim1 of the focal spot and the appropriate number of pixels Dim2 of the defocused spot according to the focal point of the laser system A, the size of the defocused spot, and the pixel size δ1 of the detector.

[0014] Preferably, in step 3), first, according to the calculation formula of the equivalent Fresnel number, and according to the equivalent Fresnel number Nf of the laser system A and the beam aperture, wavelength, and focal length of the laser system B, deduce the pre-focal distance Δz of the laser system B; then adjust the defocus position of the laser system B.

[0015] Preferably, calculate the appropriate number of pixels Dim1' of the focal spot and the appropriate number of pixels Dim2' of the defocused spot of the laser system B according to the pixel size δ2 of the detector of the laser system B, and the calculation formula is as follows:

[0016] Dim1' = Dim1 * δ1 / δ2;

[0017] Dim2' = Dim2 * δ1 / δ2;

[0018] Wherein, δ1 is the pixel size of the detector of the laser system A.

[0019] Preferably, the resampling means resampling the focus spot of Dim1’*Dim1’ collected by the laser system B into Dim1*Dim1, and resampling the defocus spot of Dim2’*Dim2’ collected by the laser system B into Dim2*Dim2 to obtain resampled data Dim1*Dim1 and Dim2*Dim2.

[0020] Preferably, in step 4), according to the obtained wavefront phase, the optical path is obtained by multiplying by λ / (2Π).

[0021] An intelligent wavefront detection parameter migration system based on the equivalent Fresnel number includes a memory and a processor. A computer program is stored on the memory, and when the processor executes the program, the steps of the above method are implemented.

[0022] The beneficial effects of the present invention are:

[0023] The technical solution of intelligent wavefront detection parameter migration based on the equivalent Fresnel number proposed by the present invention is for the intelligent wavefront detection technology based on deep learning. When the wavefront is inverted through the intensity spots at the focus and defocus positions, as long as the beam aperture shape remains unchanged, even when the wavelength, aperture, and focal length in the laser system parameters change, the wavefront reconstruction deep learning network model can perform laser system parameter migration through the calculation of the equivalent Fresnel number. By calculating the equivalent parameters corresponding to the same equivalent Fresnel number, the wavefront reconstruction deep learning network model can be directly applied in different laser systems without the need to retrain for different laser systems. Furthermore, one can focus on training a good wavefront reconstruction deep learning network model. One can combine a large number of simulation samples and a small number of experimental samples to solve the problem of difficult acquisition of actual training samples, which is convenient and fast and easy to be put into practical application. One can maximize the enrichment of the training sample library to make the wavefront reconstruction deep learning network model have strong adaptability and robustness, which is beneficial to the engineering application of intelligent wavefront detection technology. Description of the Drawings

[0024] Figure 1 is a schematic diagram of the positions of the focus and defocus double spots;

[0025] Figure 2 is a diagram of a wavefront reconstruction deep learning network model of the present invention;

[0026] Figure 3 is the optical path diagram of the laser system A in the embodiment;

[0027] Figure 4 is a schematic diagram of parameter migration applied to the laser system B in the embodiment;

[0028] Figure 5 is the experimental test result diagram. Detailed implementation manners

[0029] To make the above features and advantages of the present invention more obvious and understandable, specific embodiments are given below and described in detail in conjunction with the accompanying drawings as follows.

[0030] According to the following Fresnel diffraction transmission theory of light beams, the complex amplitude distribution of the optical field changes continuously with the transmission distance:

[0031]

[0032] In the formula, P(ξ,η) is the amplitude of the initial optical field, is the phase of the initial optical field, (ζ,η) is the coordinate of the initial optical field plane perpendicular to the optical axis, j is an imaginary number, λ is the laser wavelength, (x,y) is the coordinate of the detection optical field plane perpendicular to the optical axis, and z is the coordinate in the optical axis direction. The amplitude and phase of the optical field affect the distribution of the transmitted optical field simultaneously, thereby affecting the intensity distribution of the transmitted light.

[0033] When the optical field passes through a focusing lens, it is equivalent to adding a quadratic spherical phase, as follows:

[0034]

[0035] At the focal position, i.e., z = f, a Fraunhofer diffraction spot can be obtained, and at different defocused positions, Fresnel diffraction spots with different diffraction degrees can be obtained. This diffraction degree can be evaluated by the following equivalent Fresnel number:

[0036]

[0037] Where, Nf is the equivalent Fresnel number, a is the beam aperture, λ is the laser wavelength, f is the focal length, and Δz is the distance of the defocused position in front of the focus.

[0038] As long as the equivalent Fresnel numbers are the same, the degrees of the diffraction spots are the same. For an intelligent wavefront correction model constructed for a set of system parameters, when the shape of the beam aperture remains unchanged, for different system parameters, including different wavelengths, apertures, and focal lengths, it can be directly applied by recalculating the parameters corresponding to the same equivalent Fresnel number, without the need to retrain the intelligent wavefront network model for another set of system parameters.

[0039] Based on the above technical concept, the present invention proposes an intelligent wavefront detection parameter migration method based on the equivalent Fresnel number, and its processing steps are as follows:

[0040] Step 1: Collect training data according to the parameters of a specific laser system A, including focused and defocused spot patterns and distorted wavefront phase distributions. Use this training data to train a wavefront restoration deep learning network model (see the deep learning network in Figure 2 ), which is constructed based on CNN, CGAN, or other neural networks;

[0041] Step 2: Calculate the corresponding equivalent Fresnel number Nf according to the beam aperture, wavelength, focal length, and defocus position of laser system A. Select appropriate numbers of pixels Dim1 for the focused spot and Dim2 for the defocused spot according to the sizes of the focused and defocused spots and the pixel size δ1 of the detector;

[0042] Step 3: According to the calculation formula (3) of the equivalent Fresnel number, infer the pre-focus distance Δz for laser system B based on the equivalent Fresnel number Nf in Step 1 and the beam aperture, wavelength, and focal length of the measured laser system B. By adjusting the defocus position of laser system B, ensure that the equivalent Fresnel numbers of laser system B and laser system A are equal. Calculate the appropriate numbers of pixels Dim1’ for the focused spot and Dim2’ for the defocused spot according to the pixel size δ2 of the detector of laser system B:

[0043] Dim1’ = Dim1 * δ1 / δ2 (4)

[0044] Dim2’ = Dim2 * δ1 / δ2 (5)

[0045] Step 4: For laser system B, adjust the detection position of the defocus detector according to the defocus position calculated in Step 3, and intercept pixel sizes of Dim1’ and Dim2’ centered on the spot centroid in the focused and defocused detectors. Since the inputs of the network model are a focused spot with Dim1 * Dim1 pixels and a defocused spot with Dim2 * Dim2 pixels, resample the Dim1’ * Dim1’ focused spot and the Dim2’ * Dim2’ defocused spot collected by laser system B to Dim1 * Dim1 and Dim2 * Dim2 respectively, which can then be used as inputs to the wavefront restoration deep learning network model to restore the wavefront phase, as Figure 2 shown. The obtained wavefront phase is an absolute phase value. If it needs to be converted to optical path, it should be multiplied by λ / (2Π).

[0046] The present invention also provides an intelligent wavefront detection parameter migration system based on the equivalent Fresnel number for implementing the above method. The system includes a memory and a processor. A computer program is stored on the memory, and when the processor executes the program, the steps of the above method are implemented. The technical solution of the system can also be described as including two modules: a wavefront restoration deep learning network model and a parameter calculation module, where the parameter calculation module is used to implement the calculations in the second and third steps of the above method.

[0047] For the technical solution proposed by the present invention, a specific embodiment is given below:

[0048] For a set of laser system A, the system structure is as Figure 3 shown. Two deformable mirrors are used. The first deformable mirror is used to generate a distorted phase, and the second deformable mirror is used to correct the distorted phase. A Hartmann is used to detect the marked wavefront phase. Two CCDs are respectively used to collect the defocused and focused light spots. The AI fitting of the far-field light spot is carried out by using the light spot distribution data of CCD1 and CCD2, and the training data set is constructed by collecting the actual far-field and defocused light spots through the CCD.

[0049] The parameters of laser system A are as follows:

[0050] The light output of the laser system is a plane beam, and the aperture is circular;

[0051] The laser wavelength is 1064 nm;

[0052] The laser output aperture is 45 mm;

[0053] The focal length of the focusing lens is 1000 mm;

[0054] The defocus position is 20 mm in front of the focus;

[0055] Beam splitter prism + coating: The number is 2;

[0056] The first deformable mirror: DM1 from France, the maximum angle is 15°;

[0057] The second deformable mirror: DM2 from Tsinghua University, the maximum angle is 10°;

[0058] Aperture parameter: 45 mm × 45 mm / φ = 45 mm;

[0059] The light intensities of the focused and defocused light spots are collected through two CCDs, and the pixel size is 5 μm.

[0060] Based on the above laser system A, the method of the present invention is used for wavefront detection parameter migration, and the specific implementation steps are as follows:

[0061] Step 1: Collect training data according to the parameters of the above laser system A. The data includes the focused and defocused light spots and the distorted wavefront phase distribution, and train a wavefront restoration deep learning network model based on this training data.

[0062] The trained wavefront detection and restoration network model should follow the detailed process below during actual use. For example, it should be applied to the laser system B as shown in Figure 4 Figure. The laser wavelength of this system is 520nm, the beam aperture is 8mm, the focal length is 250mm, the CCD pixel size is 4.65um. A distorted wavefront is given through a liquid crystal light modulator, and the actual far-field and defocused light spots are collected by the CCD to reconstruct the wavefront, and then the wavefront is corrected by a deformable mirror.

[0063] Step 2: Calculate the corresponding equivalent Fresnel number Nf according to the beam aperture, wavelength, focal length, and defocus position of the above laser system A. The calculated equivalent Fresnel number is 38.8407. Select the appropriate number of pixels Dim1 = 256 for the focused light spot and Dim2 = 512 for the defocused light spot according to the sizes of the focused and defocused light spots and the pixel size 5um of the detector.

[0064] Step 3: According to the calculation formula (3) of the equivalent Fresnel number, based on the equivalent Fresnel number Nf = 38.8407 in the first step, the beam aperture 8mm, wavelength 520nm, and focal length 250mm of the measured laser system B, deduce the distance Δz = 18.3mm in front of the focus for the laser system B. By adjusting the defocus position of the laser system B, ensure that the equivalent Fresnel numbers of the laser system B and the laser system A are equal. And according to the pixel size δ2 = 4.65um of the detector of the laser system B, calculate the appropriate number of pixels Dim1’ = 189 for the focused light spot and Dim2’ = 378 for the defocused light spot.

[0065] Step 4: For the laser system B, adjust the detection position of the defocus detector according to the defocus position calculated in the third step. Intercept pixel sizes of 189 and 378 centered on the centroid of the light spot in the focused and defocus detectors, and resample to the dimension 256 for the input of the network model, which can then be used as input to the network model to restore the wavefront phase. The obtained wavefront is an absolute phase value. If it is to be converted to the optical path, it should be multiplied by λ / (2Π), where λ is 520nm.

[0066] Experimental test:

[0067] Since the liquid crystal light modulator can give a known distorted wavefront, it can be compared and verified with the wavefront detected and restored by the system. Conduct experimental tests on 200 groups of data. The experimental results are as shown in Figure 5As shown in the figure, where (a) is the actual given wavefront, PV = 2.713λ; (b) is the image given to the liquid crystal light modulator, which has a rotational relationship with (a); (c) is the focus collected by the CCD; (d) is the defocused spot collected by the CCD; (e) is the wavefront generated by the network, PV = 2.736λ; (f) is the difference between (a) and (e), root mean square = 0.113λ. From the experimental results, the far-field spots before and after correction are as follows: After testing 200 groups of data, the average root mean square error is 0.089λ, and the groups with root mean square less than λ / 10 account for 98.76% of all tested groups, verifying the effectiveness of the method of the present invention.

[0068] In this embodiment, the method of the present invention is applied in the above laser system. The network model of the intelligent wavefront system is the same and does not perform specific network model training for specific system parameters. Therefore, the method of the present invention can achieve parameter migration of the intelligent wavefront detection system based on the equivalent Fresnel number, which can promote the engineering application of the intelligent wavefront detection method.

[0069] The parts not elaborated in detail in the present invention belong to the well-known technologies of those skilled in the art.

[0070] Although the present invention has been disclosed above in embodiments, it is not intended to limit the present invention. Appropriate modifications or equivalent replacements made by those of ordinary skill in the art to the technical solutions of the present invention shall be covered within the protection scope of the present invention. The protection scope of the present invention shall be defined by the claims.

Claims

1. An intelligent wavefront detection parameter migration method based on the equivalent Fresnel number, characterized in that It includes the following steps: 1) Collect training data according to the parameters of the known laser system A, and use the training data to train a wavefront restoration deep learning network model. The training data includes the focus, defocused spots, and distorted wavefront phase distribution. The wavefront restoration deep learning network model is constructed based on CNN or CGAN; 2) Calculate the equivalent Fresnel number Nf of the laser system A according to the beam aperture, wavelength, focal length, and defocus position parameters of the laser system A, and select the appropriate number of pixels Dim1 for the focus spot and the appropriate number of pixels Dim2 for the defocused spot of the laser system A; 3) For the laser system B to be measured, adjust the defocus position of the laser system B to make the equivalent Fresnel number of the laser system B equal to that of the laser system A, and calculate the appropriate number of pixels Dim1' for the focus spot and the appropriate number of pixels Dim2' for the defocused spot of the laser system B according to the Dim1 and Dim2; 4) For the laser system B, adjust the detection position of the defocus detector according to the defocus position calculated in step 3), intercept the pixel sizes of Dim1' and Dim2' centered on the spot centroid in the focus and defocus detectors, perform resampling, and then input the resampled data into the wavefront restoration deep learning network model to obtain the wavefront phase of the laser system B.

2. The method according to claim 1, wherein Select the appropriate number of pixels Dim1 for the focus spot and the appropriate number of pixels Dim2 for the defocused spot of the laser system A according to the focus, defocused spot size of the laser system A, and the pixel size δ1 of the detector.

3. The method according to claim 1, wherein In step 3), first, according to the calculation formula of the equivalent Fresnel number, calculate the pre-focal distance ∆z of the laser system B according to the equivalent Fresnel number Nf of the laser system A and the beam aperture, wavelength, and focal length of the laser system B; then adjust the defocus position of the laser system B.

4. The method according to claim 1, wherein Calculate the appropriate number of pixels Dim1' for the focus spot and the appropriate number of pixels Dim2' for the defocused spot of the laser system B according to the pixel size δ2 of the detector of the laser system B. The calculation formulas are as follows: Dim1' = Dim1 * δ1 / δ2; Dim2' = Dim2 * δ1 / δ2; where δ1 is the pixel size of the detector of the laser system A.

5. An intelligent wavefront detection parameter migration system based on the equivalent Fresnel number, characterized in that, It includes a memory and a processor. A computer program is stored on the memory. When the processor executes the program, it implements the steps of the method described in any one of claims 1-4.

Images