Aberration measurement method based on liquid crystal geometric phase plate and deep learning

Abstract

The invention discloses an aberration measurement method based on a liquid crystal geometric phase plate and deep learning, and belongs to the technical field of photoelectric measurement. According to the invention, the constructed computer-generated hologram is processed into the liquid crystal geometric phase plate. And the computer-generated hologram encodes the plurality of Lukosz aberration modes in a phase encoding mode. A plurality of phase offsets are applied to an incident light wave. And sub-light-spot segmentation extraction is carried out on the collected light spot array image, and all sub-light spots are combined into a three-dimensional light spot image with a plurality of channels. And constructing a deep learning network model for wavefront aberration measurement. In the network training stage, a spatial light modulator is adopted to generate known aberration, and a three-dimensional light spot image is used as an input feature sequence of a network model. And training the network model by using the training data set. Wavefront modulation is carried out through the liquid crystal geometric phase plate, simultaneous acquisition of a plurality of offset images is realized, and the image acquisition efficiency is improved. A deep learning network is used to train and predict a light spot image, and high-precision real-time measurement of aberration is realized.

Description

Technical Field

[0001] The present invention relates to an aberration measurement method based on a liquid crystal geometric phase plate and deep learning, belonging to the technical field of optoelectronic measurement. Background Art

[0002] When a spherical wavefront emitted from a point on an object propagates, it may be affected by various error sources and generate wavefront aberration. Common error sources include non-uniformity of the transmission medium, machining and alignment errors of optical elements, etc. Aberration distorts the shape of the wavefront and no longer focuses on a single point, resulting in the inability of the optical system to clearly image the object.

[0003] For an imaging optical system, the measurement accuracy of wavefront aberration directly affects image quality such as image resolution and contrast. During the design, machining, and alignment stages of the optical system, wavefront aberration needs to be measured. Currently, common wavefront aberration measurement methods mainly include interference measurement method, Shack-Hartmann wavefront sensing method, holographic mode wavefront sensing method, phase retrieval method, etc.

[0004] The interference measurement method interferes the measured light wave with a reference light wave. The generated fringe image records the aberration information of the measured wavefront, and the wavefront aberration is extracted by analyzing the fringes. The interference measurement method has high accuracy, but it requires the introduction of a reference light wave, the system structure is relatively complex, and it is easily affected by environmental factors such as mechanical vibration.

[0005] The Shack-Hartmann wavefront sensor uses a microlens array to divide the measured wavefront into multiple sub-regions. Each microlens focuses the wavefront of the sub-region. By measuring the centroid offset of the focused spot in each sub-region, the wavefront gradient information can be measured, and then the wavefront aberration can be obtained through the wavefront reconstruction method. The Shack-Hartmann wavefront sensor has the advantages of fast measurement speed, large dynamic range, and large spectral range, but is limited by the microlens processing technology and has low spatial resolution.

[0006] The holographic mode wavefront sensor uses a computer-generated hologram to modulate the incident wavefront. By measuring the evaluation function of the modulated image, such as the intensity at the center of the spot, the second moment of the spot, etc., the mode coefficients corresponding to the wavefront aberration can be calculated. The advantages of the holographic mode wavefront sensor are simple calculation and fast speed, but the measurement accuracy is not high and the spectral range is narrow.

[0007] The phase retrieval method calculates the wavefront aberration through iterative operations or non-linear optimization, and has the advantages of simple structure and high measurement accuracy. However, it needs to introduce multiple phase differences, the image acquisition time is long, the algorithm iteration time is long, and it is easy to fall into local extrema. Summary of the Invention

[0008] The object of the present invention is to provide a wavefront aberration measurement method based on a liquid crystal geometric phase plate and deep learning. By constructing a multi-aberration mode multiplexing computer-generated hologram and processing it into a liquid crystal geometric phase plate, the liquid crystal geometric phase plate is used to modulate the incident light wave, so that multiple offset images are simultaneously presented on the image plane. A deep learning network is built to extract features from the offset images, realizing a non-linear mapping from the offset images to the aberration mode coefficients, and achieving the purpose of synchronous measurement of multiple aberration modes. The present invention has the advantages of simple and compact structure, fast measurement speed and high measurement accuracy.

[0009] The object of the present invention is achieved by the following technical solutions:

[0010] A wavefront aberration measurement method based on a liquid crystal geometric phase plate and deep learning disclosed by the present invention. The optical devices for acquiring images include a circular polarizer, a liquid crystal geometric phase plate, a Fourier lens, and a planar array detector. After the incident light wave passes through the optical imaging system, a circular polarizer and a liquid crystal geometric phase plate are successively placed at the position of the secondary exit pupil of the optical imaging system. The circular polarizer modulates the polarization state of the incident light wave into a circular polarization state. The circularly polarized light is incident on the liquid crystal geometric phase plate, and the transmitted light is converted into circularly polarized light with the opposite rotation direction and superimposed with a geometric phase. The magnitude of the geometric phase is determined by the optical axis angle of the liquid crystal molecules. Through the laser direct writing processing method, the optical axis angle of each liquid crystal molecule on the liquid crystal geometric phase plate is precisely controlled, thereby realizing the geometric phase regulation of the transmitted light wave. The present invention processes the constructed computer-generated hologram into a liquid crystal geometric phase plate. The computer-generated hologram encodes multiple Lukosz aberration modes through phase encoding, and each Lukosz aberration mode is superimposed with a digital blazed grating with a different spatial frequency. The liquid crystal geometric phase plate processed according to the constructed computer-generated hologram can apply multiple phase biases to the incident light wave. After passing through the Fourier lens, the light wave is split into several sub-light waves propagating in different directions under the action of the digital blazed grating, and each sub-light wave carries a specific phase bias. All the sub-light waves are focused into a spot array image on the rear focal plane of the Fourier lens. The sub-spot segmentation extraction is performed on the collected spot array image, and all the sub-spots are combined into a three-dimensional spot image with multiple channels. A deep learning network model for wavefront aberration measurement is constructed. The input of the network model is the three-dimensional spot image, and the output is the aberration mode coefficient. In the network training stage, a spatial light modulator is used to generate known aberrations. The three-dimensional spot image is used as the input feature sequence of the network model, and the aberration mode coefficient corresponding to the known aberration is used as the label to construct a training data set. The training data set is used to train the network model. The trained network directly predicts the aberration mode coefficient contained in the incident light wave according to the input three-dimensional spot image. Through the wavefront modulation by the liquid crystal geometric phase plate, the simultaneous acquisition of multiple offset images is realized, the structure of the optical system is simplified, and the image acquisition efficiency is improved. The deep learning network is used to train and predict the spot image to realize the high-precision real-time measurement of the aberration, and the accuracy and efficiency of the wavefront aberration correction of the optical system are improved according to the aberration measurement result.

[0011] An aberration measurement method based on a liquid crystal geometric phase plate and deep learning disclosed by the present invention includes the following steps:

[0012] Step 1: Construct a multi-mode multiplexed computer-generated hologram to encode multiple Lukosz aberration modes through phase encoding, and each Lukosz aberration mode is superimposed with a digital blazed grating with a different spatial frequency.

[0013] Construct a multi-mode multiplexed computer-generated hologram The computer-generated hologram contains N - order Lukosz modes, as shown in Equation (1):

[0014]

[0015] where x represents the spatial coordinate vector, a n is the amplitude coefficient, which is used to control the energy proportion of the sub - light waves; b n is the bias coefficient, which is used to control the size of the bias mode; j is the imaginary unit, Ln(x) is the Lukosz mode, kn is the wave vector of the sub - light wave, and kn·x constitutes a digital blazed grating, which is used to control the position of the light spot on the focal plane; arg{} represents taking the argument.

[0016] Step 2: The liquid - crystal geometric phase plate fabricated according to the constructed computer - generated hologram can apply multiple phase biases to the incident light wave. After passing through the Fourier lens, the light wave is split into several sub - light waves propagating in different directions under the action of the digital blazed grating, and each sub - light wave carries a specific phase bias. All the sub - light waves are focused into a spot array image on the rear focal plane of the Fourier lens. The sub - spots in the collected spot array image are segmented and extracted, and all the sub - spots are combined into a three - dimensional spot image with multiple channels. The liquid - crystal molecules are equivalent to a wave plate with an adjustable optical axis angle θ.

[0017] The Jones matrix J corresponding to the liquid - crystal molecules is:[[]]

[0018]

[0019] In the formula, δ = 2πdΔn / λ, where δ represents the phase delay of the liquid - crystal molecules; d represents the thickness of the liquid - crystal layer; Δn represents the birefringence of the liquid - crystal molecules; λ represents the wavelength of the incident light wave; R(θ) represents the spatial coordinate rotation.

[0020] The Jones vector of left - hand circularly polarized light is The Jones vector of right - hand circularly polarized light is When a left - hand or right - hand circularly polarized light is incident on the liquid - crystal molecules, the polarization vector of the transmitted light is:[[]]

[0021]

[0022] Controlling the thickness d of the liquid - crystal molecule layer can make the phase delay δ of the liquid - crystal molecules satisfy the half - wave condition, that is, δ=(2k +

[0023] 1)π, where k is an integer. Substituting into formula (3), the first term of formula (3) can be eliminated. The polarization direction of the transmitted light is opposite to that of the incident light and carries a geometric phase with a magnitude of 2θ. By the laser direct writing method, the optical axis angle of the liquid - crystal molecules is set to Then, the computer-generated hologram constructed in Step 1 is processed into a liquid crystal geometric phase plate.

[0024] Step 3: Set up a measurement optical path for obtaining images. The measurement optical path includes a circular polarizer, a liquid crystal geometric phase plate, a Fourier lens, and a planar array detector. After the incident light wave passes through the measurement optical path, the circular polarizer and the liquid crystal geometric phase plate are successively placed at the position of the secondary exit pupil of the optical imaging system. The circular polarizer modulates the polarization state of the incident light wave into a circular polarization state. The circularly polarized light is incident on the liquid crystal geometric phase plate, and the transmitted light is converted into circularly polarized light with the opposite rotation direction and superimposed with a geometric phase. The magnitude of the geometric phase is determined by the optical axis angle of the liquid crystal molecules. By means of the laser direct writing processing method, the optical axis angle of each liquid crystal molecule on the liquid crystal geometric phase plate is precisely controlled, thereby realizing the geometric phase regulation of the transmitted light wave. The circularly polarized light is incident on and passes through the liquid crystal geometric phase plate, and the polarization direction of the outgoing light is opposite to that of the incident light, and a geometric phase with a magnitude of is superimposed.

[0025] Step 4: Collect images. The light wave modulated by the liquid crystal geometric phase plate is split into N sub-light waves under the action of the digital blazed grating. Each sub-light wave is focused on different spatial positions on the planar array detector under the combined action of the wave vector kn and the Fourier lens. The focused light spots form a light spot array image, which is collected by the planar array detector. The collected light spot array image is cropped and segmented to extract all sub-light spot images, and combined into a three-dimensional light spot image with N channels and a resolution of m1×m2.

[0026] Step 5: Set up a deep learning network model. The deep learning network model is mainly composed of convolutional layers, max pooling layers, and fully connected layers. A batch normalization layer is added after each convolutional layer, and at the same time, the CBAM attention mechanism is added to improve the performance and generalization ability of the deep learning network model. A Dropout layer is added to the deep learning network model to prevent the problem of overfitting of the network. The input of the deep learning network model is the three-dimensional light spot image with N channels collected in Step 4. The convolutional layers are used to extract image features at different levels in the image. The pooling layer is applied after the convolutional layer to reduce the spatial size of the feature map and reduce the number of parameters. The pooling layer provides translational invariance for the model, which helps the network model to identify features independent of the position in the image. The fully connected layer is used at the end of the network model. Each unit in the fully connected layer is tightly connected to all neurons in the previous layer, and integrates and outputs the image features extracted by all the previous convolutional layers. The output of the network model is the N Lukosz aberration mode coefficients.

[0027] Step 6: Train the deep learning network model. Configure the parameters required for training the deep learning network model, set the learning rate, batch size, weight initialization method, optimization method, and number of iterations. Generate known wavefront aberrations through a spatial light modulator, fit to obtain N Lukosz aberration mode coefficients, which are used as labels. The three-dimensional spot image and the label constitute the training dataset, and a part is randomly selected from the training dataset according to a ratio as the validation set. Input the three-dimensional spot image collected and processed by the area array detector into the network model. Take the mean square error between the N values output by the network and the label values as the loss function (Loss). Use the gradient backpropagation algorithm and the Adam optimizer to iteratively optimize the network parameters. During the training process, monitor the Loss curve of the validation set in real time. When the Loss value of the validation set no longer decreases within a predetermined number of times, end the training. Save the parameters of the trained deep learning network model.

[0028] Step 7: Use the trained deep learning network model to measure wavefront aberrations. Input the collected three-dimensional spot image into the trained network model to obtain N Lukosz aberration mode coefficients q1 to q N , linearly combine the aberration mode coefficients with the corresponding Lukosz aberration modes, to obtain the two-dimensional wavefront map Φ corresponding to the wavefront aberration, that is, realize aberration measurement based on the liquid crystal geometric phase plate and deep learning.

[0029] Beneficial effects:

[0030] 1. For the aberration measurement method based on the liquid crystal geometric phase plate and deep learning disclosed in the present invention, the optical devices for acquiring images include a circular polarizer, a liquid crystal geometric phase plate, a Fourier lens, and an area array detector. After the incident light wave passes through the optical imaging system, a circular polarizer and a liquid crystal geometric phase plate are successively placed at the secondary exit pupil position of the optical imaging system. The circular polarizer modulates the polarization state of the incident light wave into a circular polarization state. The circularly polarized light is incident on the liquid crystal geometric phase plate, and the transmitted light is converted into circularly polarized light with the opposite rotation direction and superposed with a geometric phase. The magnitude of the geometric phase is determined by the optical axis angle of the liquid crystal molecules. By using the laser direct writing processing method, the optical axis angle of each liquid crystal molecule on the liquid crystal geometric phase plate is precisely controlled, thereby realizing the geometric phase regulation of the transmitted light wave. At the same time, multiple spot images with phase biases are collected to avoid errors caused by multiple movements of the components, and the image acquisition efficiency is high.

[0031] 2. The aberration measurement method based on a liquid crystal geometric phase plate and deep learning disclosed by the present invention processes the constructed computer-generated hologram into a liquid crystal geometric phase plate. The computer-generated hologram encodes multiple Lukosz aberration modes through phase encoding, and each Lukosz aberration mode is superimposed with a digital blazed grating of a different spatial frequency. The liquid crystal geometric phase plate processed according to the constructed computer-generated hologram can apply multiple phase biases to the incident light wave. After passing through a Fourier lens, the light wave is split into several sub-light waves propagating in different directions under the action of the digital blazed grating, and each sub-light wave carries a specific phase bias. All sub-light waves are focused into a spot array image on the rear focal plane of the Fourier lens. The sub-spots in the collected spot array image are segmented and extracted, and all sub-spots are combined into a three-dimensional spot image with multiple channels. A deep learning network model for wavefront aberration measurement is constructed. By deep learning the deep features of the spot image, the non-linear relationship between the spot image and the Lukosz aberration coefficients is fitted, which has the advantage of fast measurement speed.

[0032] 3. The aberration measurement method based on a liquid crystal geometric phase plate and deep learning disclosed by the present invention uses the image detector of the system to be measured itself as the image acquisition device, without the need to equip a dedicated detector, and the measurement cost is low. The liquid crystal geometric phase plate is small in volume, the liquid crystal processing technology is mature, and the processing cost is low. When in use, there is no need to split the incident light wave, avoiding the non-common path error. Multiple bias images are collected simultaneously, avoiding the measurement error easily caused during the process of switching devices, and having the advantages of simple and compact structure and strong stability.

[0033] 4. The aberration measurement method based on a liquid crystal geometric phase plate and deep learning disclosed by the present invention uses a training data set to train the network model. The trained network directly predicts the aberration mode coefficients contained in the incident light wave according to the input three-dimensional spot image. Through wavefront modulation by the liquid crystal geometric phase plate, multiple bias images are collected simultaneously. The output aberration mode type and the number of aberration modes of the deep learning network can be flexibly adjusted according to the usage scenario, having the advantage of wide application range. Using the deep learning network to train and predict the spot image, high-precision real-time measurement of aberration is realized, and the accuracy and efficiency of wavefront aberration correction of the optical system are improved according to the aberration measurement results. Description of the Drawings

[0034] Figure 1 In (a) is a schematic diagram of spot image acquisition; Figure 1 In (b) is a schematic diagram of network model training; Figure 1 In (c) is a schematic diagram of network model testing.

[0035] Figure 2 is the structural diagram of the constructed network model.

[0036] Figure 3 Schematic diagrams of the designed computer-generated hologram and liquid crystal geometric phase plate.

[0037] Figure 4 For the measured value and measurement error of an input aberration.

[0038] Figure 5 Schematic diagram of the optical path structure for data set acquisition.

[0039] Figure 6 Flowchart of the aberration measurement method based on liquid crystal geometric phase plate and deep learning of the present invention. Detailed implementation mode

[0040] To better illustrate the purpose and advantages of the present invention, the following further describes the invention content in conjunction with the drawings and examples.

[0041] Example 1:

[0042] As Figure 6 shown, the aberration measurement method based on liquid crystal geometric phase plate and deep learning disclosed in this example is specifically implemented as follows:

[0043] Step 1: Design a multi-mode multiplexing computer-generated hologram The computer-generated hologram includes 15th-order Lukosz modes, as shown in Equation (1):

[0044]

[0045] Among them, the amplitude coefficient a n is used to control the energy proportion of the sub-light waves, and the specific settings are shown in Table 1; the bias coefficient b n is used to control the size of the bias mode, and the specific settings are shown in Table 2; the wave vector k of the sub-light wave n is used to control the position of the light spot on the focal plane, and the specific settings are shown in Table 3. The designed computer-generated hologram is as shown in the appendix Figure 3 shown.

[0046] Table 1. Amplitude coefficient of the computer-generated hologram

[0047] <![CDATA[a4]]> <![CDATA[a5]]> <![CDATA[a6]]> <![CDATA[a7]]> <![CDATA[a8]]> <![CDATA[a9]]> <![CDATA[a 10 > <![CDATA[a 11 > <![CDATA[a 12 > <![CDATA[a 13 > <![CDATA[a 14 > <![CDATA[a 15 > <![CDATA[a 16 > <![CDATA[a 17 > <![CDATA[a 18 > <![CDATA[a 19 > 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0.5

[0048] Table 2. Bias coefficient of the computer-generated hologram

[0049] <![CDATA[b4]]> <![CDATA[b5]]> <![CDATA[b6]]> <![CDATA[b7]]> <![CDATA[b8]]> <![CDATA[b9]]> <![CDATA[b 10 > <![CDATA[b 11 > <![CDATA[b 12 > <![CDATA[b 13 > <![CDATA[b 14 > <![CDATA[b 15 > <![CDATA[b 16 > <![CDATA[b 17 > <![CDATA[b 18 > <![CDATA[b 19 > 20 20 20 20 20 20 20 20 20 20 20 20 20 20 20 0

[0050] Table 3. Wave vector of the computer-generated hologram

[0051]

[0052] Step 2: Process the designed computer-generated hologram into a liquid crystal geometric phase plate. The liquid crystal geometric phase plate is processed from liquid crystal polymer materials. Liquid crystal molecules can be equivalent to a wave plate with an adjustable optical axis angle, and the phase delay δ = 2πdΔn / λ. λ represents the wavelength of the incident light wave, designed to be 635 nm; Δn represents the birefringence of liquid crystal molecules, which is related to the specific liquid crystal material; d represents the thickness of the liquid crystal layer; control Δn and d to ensure that the liquid crystal molecules meet the half-wave condition, that is, δ = (2k + 1)π, where k is an integer. Use laser direct writing processing technology to control the optical axis angle of the liquid crystal molecules to meet Process the computer-generated hologram designed in Step 1 into a liquid crystal geometric phase plate. The schematic diagram of the liquid crystal geometric phase plate is shown in the appendix Figure 3 as follows.

[0053] Step 3: Set up the measurement optical path as shown in the appendix Figure 5 as follows. The spherical wave emitted by the single-mode fiber laser source is collimated by the lens L1, and after passing through the polarizer P, it is modulated into a linearly polarized light. The linearly polarized light passes through the beam splitter prism BS and is incident on the spatial light modulator SLM. The spatial light modulator generates wavefront aberrations. The reflected light carries the aberrations and passes through the beam splitter prism again. After reflection, it is focused by the lens L2 onto the pinhole. The pinhole filters out the orders other than the zero order. The zero-order diffracted light passing through the pinhole is collimated again by the lens L3. It passes through a quarter-wave plate in sequence and is modulated into a circularly polarized light, and then passes through the liquid crystal geometric phase plate to generate geometric phases. Finally, it is focused and imaged onto the area array detector by the lens L4.

[0054] Step 4: Collect the data set. Use the spatial light modulator to generate 20,000 random wavefront aberrations, and use the 15 Lukosz mode coefficients corresponding to the wavefront aberrations as labels. Use the area array detector to collect 20,000 corresponding spot array images. After segmentation and extraction in the manner shown in appendix Figure 1 (a), the resolution of each sub-image is 128×128 pixels. Combine all the spots in the spot array image into an image with 16 channels. Finally, obtain a data set of 20,000 spot images of 16×128×128. Randomly select 2,000 images and the corresponding labels as the validation set. Additionally, independently generate 1,000 groups of aberrations and spot images as the test set.

[0055] Step 5: Set up a deep learning network model. As shown in the appendix Figure 2As shown in the figure, the network mainly consists of 5 convolutional layers, 5 max pooling layers, and 1 fully connected layer. A batch normalization layer is added after each convolutional layer, and the CBAM attention mechanism is added after the 2nd and 4th pooling layers to improve the performance and generalization ability of the model. The input of the network model is the three-dimensional spot image with a dimension of 16×128×128 collected in Step 3. The convolutional kernel size of the convolutional layer is 3×3, which is used to extract image features at different levels in the image. The pooling layer is applied after the convolutional layer, and the pooling kernel size is 2×2, which halves the spatial size of the feature map and reduces the number of parameters. The pooling layer provides translational invariance for the model and helps the network model identify features that are independent of the position in the image. The fully connected layer is used at the end of the network model. Each unit in the fully connected layer is tightly connected to all neurons in the previous layer, and integrates and outputs the image features extracted by all the previous convolutional layers. The output of the network model is 15 Lukosz aberration mode coefficients.

[0056] Step 6: Train the network model. Configure the parameters required for network model training, set the learning rate to 0.001, the learning rate decay to 0.001, the batch size to 50, the optimizer to Adam, and the number of training epochs to 200. Input the three-dimensional spot image collected and processed by the area array detector into the network model. Take the mean square error between the 15 coefficient values output by the network and the label values as the loss function (Loss), as shown in Appendix Figure 1 (b). Use the gradient backpropagation algorithm and the Adam optimizer to iteratively optimize the network parameters. During the training process, monitor the Loss curve of the validation set in real time. When the Loss value of the validation set does not decrease within 20 times, stop the training. Save the parameters of the trained network model.

[0057] Step 7: Use the trained network model to measure the wavefront aberration. The input aberration coefficients and the aberration coefficients measured by the network model are shown in Table 4. The two-dimensional wavefront diagrams corresponding to the input aberration and the network output aberration are shown in Appendix Figure 4 As shown, the RMS of the input wavefront aberration is 0.5 wavelength, the RMS of the wavefront aberration measured by the network is 0.51 wavelength, and the RMS value of the measurement error is 0.02 wavelength.

[0058] Table 4. Input and measured aberration coefficients

[0059] Mode order 4 5 6 7 8 9 10 11 Input value 6.3 2.5 -4.2 -4.0 -5.1 -2.9 4.9 1.6 Measured value 6.4 2.4 -4.3 -4.2 -5.1 -2.8 4.8 1.5 Mode order 12 13 14 15 16 17 18 Input value 0.3 -4.8 1.3 0.1 3.5 -1.7 1.2 Measured value 0.2 -4.8 1.5 0.1 3.9 -1.5 1.1

[0060] An aberration measurement method based on a liquid crystal geometric phase plate and deep learning disclosed in this embodiment processes the designed multi-mode multiplexed computer-generated hologram into a liquid crystal geometric phase plate to modulate the incident light wave, and simultaneously collects multiple spot images with introduced phase biases, which can effectively improve the image acquisition efficiency and simplify the optical path structure. The type and number of aberration modes to be measured can be flexibly adjusted, and it has the advantages of a wide application scenario range, high measurement accuracy, and fast measurement speed.

[0061] The specific description above further details the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above is only a specific embodiment of the present invention and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for measuring aberrations based on a liquid crystal geometric phase plate and deep learning, characterized in that: It includes the following steps: Step 1: Construct a multi-mode multiplexed computer-generated hologram (CGH). Encode multiple Lukosz aberration modes by phase encoding, and superimpose a digital blazed grating with a different spatial frequency on each Lukosz aberration mode; Step 2: The liquid crystal geometric phase plate fabricated according to the constructed CGH can apply multiple phase biases to the incident light wave. After passing through the Fourier lens, the light wave is split into several sub-light waves propagating in different directions under the action of the digital blazed grating, and each sub-light wave carries a specific phase bias; all sub-light waves are focused into a spot array image on the back focal plane of the Fourier lens; perform sub-spot segmentation extraction on the collected spot array image, and combine all sub-spots into a three-dimensional spot image with multiple channels; the liquid crystal molecules are equivalent to a wave plate with an adjustable optical axis angle; Step 3: Build a measurement optical path for acquiring images. The measurement optical path includes a circular polarizer, a liquid crystal geometric phase plate, a Fourier lens, and a planar array detector; after the incident light wave passes through the measurement optical path, place a circular polarizer and a liquid crystal geometric phase plate successively at the position of the secondary exit pupil of the optical imaging system; The circular polarizer modulates the polarization state of the incident light wave into a circular polarization state. The circularly polarized light is incident on the liquid crystal geometric phase plate, and the transmitted light is converted into circularly polarized light with the opposite sense of rotation and superimposed with a geometric phase. The magnitude of the geometric phase is determined by the optical axis angle of the liquid crystal molecules. By means of the laser direct writing processing method, the optical axis angle of each liquid crystal molecule on the liquid crystal geometric phase plate is precisely controlled, thereby realizing the geometric phase regulation of the transmitted light wave. When the circularly polarized light is incident on the liquid crystal geometric phase plate, the polarization direction of the outgoing light is opposite to that of the incident light, and a geometric phase with a magnitude of is superimposed; Step 4: Image acquisition; The light wave modulated by the liquid crystal geometric phase plate is split into N sub-light waves under the action of the computer-generated hologram. Each sub-light wave is focused on different spatial positions on the area array detector under the combined action of the wave vector k n and the Fourier lens, and the focused light spots form a light spot array image, which is acquired by the area array detector; The acquired light spot array image is cropped and segmented to extract all sub-light spot images, and combined into a three-dimensional light spot image with N channels and a resolution of m1×m2; Step 5: Build a deep learning network model; the deep learning network model is mainly composed of convolutional layers, max pooling layers, and fully connected layers; add a batch normalization layer after each convolutional layer, and at the same time add a CBAM attention mechanism to improve the performance and generalization ability of the deep learning network model; add a Dropout layer to the deep learning network model to prevent overfitting of the network; the input of the deep learning network model is the three-dimensional spot image with N channels collected in Step 2; the convolutional layer is used to extract image features at different levels in the image; the pooling layer is applied after the convolutional layer to reduce the spatial size of the feature map and reduce the number of parameters; the pooling layer provides translational invariance for the model, which helps the network model to identify features independent of the position in the image; the fully connected layer is used at the end of the network model, and each unit in the fully connected layer is tightly connected to all neurons in the previous layer to integrate and output the image features extracted by all previous convolutional layers; the output of the network model is the coefficients of N Lukosz aberration modes; Step 6: Train the deep learning network model; configure the parameters required for training the deep learning network model, set the learning rate, batch size, weight initialization method, optimization method, and number of iterations; Generate known wavefront aberrations through a spatial light modulator, fit to obtain the coefficients of N Lukosz aberration modes as labels; the three-dimensional spot image and the labels constitute the training data set, and randomly select a part from the training data set according to a ratio as the validation set; input the three-dimensional spot image collected and processed by the planar array detector into the network model; use the mean square error between the N values output by the network and the label values as the loss function (Loss); adopt the gradient backpropagation algorithm and use the Adam optimizer to iteratively optimize the network parameters; during the training process, monitor the Loss curve of the validation set in real time, and end the training when the Loss value of the validation set no longer decreases within a predetermined number of times; Save the parameters of the trained deep learning network model; Step 7: Perform wavefront aberration measurement using the trained deep learning network model; input the acquired three-dimensional spot image into the trained network model to obtain N Lukosz aberration mode coefficients q1 to q N , linearly combine the aberration mode coefficients with the corresponding Lukosz aberration modes, to obtain the two-dimensional wavefront map Φ corresponding to the wavefront aberration, that is, realize aberration measurement based on the liquid crystal geometric phase plate and deep learning.

2. The method for measuring aberration based on a liquid crystal geometric phase plate and deep learning according to claim 1, wherein: In Step 1, Constructing a Multi-Mode Reusable Computer-Generated Hologram The computer-generated hologram contains Nth-order Lukosz modes, as shown in Equation (1): where x represents the spatial coordinate vector, a n is the amplitude coefficient used to control the energy proportion of the sub-light wave; b n is the bias coefficient used to control the size of the bias pattern; j is the imaginary unit, and L n (x) is the Lukosz pattern, and k n is the wave vector of the sub-light wave. k n ·x constitutes a digital blazed grating used to control the position of the light spot on the focal plane; arg{} represents taking the argument.

3. The method for measuring aberration based on a liquid crystal geometric phase plate and deep learning according to claim 2, wherein: In Step 2, The Jones matrix J corresponding to the liquid crystal molecules is: where δ = 2πdΔn / λ, δ represents the phase delay of the liquid crystal molecules; d represents the thickness of the liquid crystal layer; Δn represents the birefringence of the liquid crystal molecules; λ represents the wavelength of the incident light wave; R(θ) represents the rotation of the spatial coordinates. The Jones vector of left-handed circularly polarized light is The Jones vector of right-handed circularly polarized light is When a left-handed or right-handed circularly polarized light is incident on liquid crystal molecules, the polarization vector of the transmitted light is: Controlling the thickness d of the liquid crystal molecular layer enables the phase delay amount δ of the liquid crystal molecules to satisfy the half-wave condition, that is, δ=(2k + 1)π, where k is an integer; substituting into formula (3), the first term of formula (3) is eliminated; the polarization direction of the transmitted light is opposite to that of the incident light and carries a geometric phase with a magnitude of 2θ; the optical axis angle of the liquid crystal molecules is set by the laser direct writing method to the computational hologram constructed in step one is processed into a liquid crystal geometric phase plate.

Images