Strong flicker perception wavefront sensor wavefront recovery inference method and related products
Patent Information
- Application Number
- CN202611213023.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-09-11
AI Technical Summary
[0003]现有的波前校正方式普遍利用波前传感器采集相位信息开展波前校正，强闪烁条件下的波前传感器容易出现探测失效的情况，且部分无波前传感方案依赖相位真值标注或纯仿真监督训练，缺少实测光场传播约束，波前反演精度有限，难以有效实施共轭校正，最终导致自由空间光通信接收耦合效率降低
[0016]从以上技术方案可以看出，相较于现有技术，本申请具有以下优点：
Smart Images
Figure CN122740907A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of free-space optical communication technology, and in particular to a wavefront recovery inference method and related products for strong flicker sensing without wavefront sensors. Background Technology
[0002] Atmospheric turbulence is a common phenomenon in free-space optical communication transmission links. Atmospheric turbulence can cause wavefront distortion and strong scintillation, which directly affects the quality of the optical field at the receiving end. Wavefront correction is usually required to compensate for the distortion in order to ensure communication reception performance.
[0003] Existing wavefront correction methods generally use wavefront sensors to collect phase information for wavefront correction. However, wavefront sensors are prone to detection failure under strong scintillation conditions. Furthermore, some wavefront-sensorless schemes rely on phase truth labeling or pure simulation-supervised training, lacking actual measured optical field propagation constraints. This results in limited wavefront inversion accuracy and difficulty in effectively implementing conjugate correction, ultimately leading to reduced free-space optical communication receiving coupling efficiency.
[0004] Therefore, how to improve the receiving coupling efficiency of free-space optical communication is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] To address the aforementioned issues, this application provides a wavefront recovery inference method and related products for strong flicker sensing without a wavefront sensor. The wavefront prediction model is trained through a dual-plane light intensity self-supervised method, eliminating the need for an external wavefront sensor. Wavefront recovery and conjugate correction are achieved solely using a single-plane light intensity distribution, making it suitable for atmospheric turbulence and strong flicker scenarios and improving the receiving coupling efficiency of free-space optical communication.
[0006] In a first aspect, embodiments of this application provide a wavefront recovery inference method for strong flicker sensing without a wavefront sensor, comprising: Acquire the first target light intensity distribution of the target light wave on the first measurement plane of the receiver; The first target light intensity distribution is input into the target wavefront prediction model to obtain the corresponding target estimated phase; the target wavefront prediction model is trained offline through a dual-plane light intensity self-supervised method; the dual-plane light intensity self-supervised method uses two sets of plane measured light intensity distributions with different transmission distances to construct the light field propagation consistency loss, iteratively optimizes the model parameters, and does not require an external wavefront sensor to provide phase labeling; The conjugate correction phase is determined based on the target estimated phase, and the target light wave is corrected using the conjugate correction phase.
[0007] Optionally, the dual-plane light intensity self-monitoring method includes: The first training light intensity distribution of the training light wave is collected on the first measurement plane, and the second training light intensity distribution of the training light wave is collected simultaneously on the second measurement plane; the transmission distance between the first measurement plane and the second measurement plane is calibrated. The first training light intensity distribution is input into the initial wavefront prediction model to obtain the training estimated phase; A first estimated complex optical field is constructed by combining the first measured amplitude corresponding to the first training light intensity distribution and the estimated phase of the training light field. The first estimated complex optical field is then transmitted to the second measurement plane through the Angular Spectrum Method (ASM) model to obtain the second estimated complex optical field. The parameters of the initial wavefront prediction model are updated based on the photometric consistency loss between the estimated propagation amplitude corresponding to the second estimated complex optical field and the second measured amplitude corresponding to the second training light intensity distribution.
[0008] Optionally, the photometric consistency loss is the L2 norm loss between the estimated propagation amplitude and the second measured amplitude.
[0009] Optionally, the target wavefront prediction model is a Zernike-Modulated Swin-Unet (ZeMS-Net).
[0010] Optionally, the ZeMS-Net includes: a Shifted Window Transformer (Swin-Transformer) encoder, a Zernike prediction branch, a residual decoding branch, and a Feature-wise Linear Modulation (FiLM) module; During offline training, a stopping gradient operation is performed on the conditional vector input to the FiLM module.
[0011] Optionally, the acquisition of the first training light intensity distribution of the training light wave on the first measurement plane, and the simultaneous acquisition of the second training light intensity distribution of the training light wave on the second measurement plane, includes: The training light wave is split to obtain a first beam and a second beam; the first beam is transmitted along a first optical path to a first measurement plane, and the second beam is transmitted along a second optical path to a second measurement plane; The first training light intensity distribution corresponding to the training light wave is acquired in the first measurement plane; The second training light intensity distribution corresponding to the training light wave is acquired in the second measurement plane.
[0012] Optionally, the step of correcting the target light wave using the conjugate correction phase includes: The conjugate correction phase is applied to a spatial light modulator or deformable mirror to correct the target light wave.
[0013] Secondly, embodiments of this application provide a wavefront recovery inference system for strong flicker sensing without a wavefront sensor, comprising: The acquisition module is used to acquire the first target light intensity distribution of the target light wave on the first measurement plane of the receiving end; The prediction module is used to input the first target light intensity distribution into the target wavefront prediction model and obtain the corresponding target estimated phase. The target wavefront prediction model is trained offline through a dual-plane light intensity self-supervised method. The dual-plane light intensity self-supervised method uses two sets of plane measured light intensity distributions with different transmission distances to construct the light field propagation consistency loss, iteratively optimizes the model parameters, and does not require an external wavefront sensor to provide phase labeling. The correction module is used to determine a conjugate correction phase based on the target estimated phase, and to correct the target light wave using the conjugate correction phase.
[0014] Thirdly, embodiments of this application provide a wavefront recovery inference device for strong flicker sensing without a wavefront sensor, comprising: Memory, used to store computer programs; A processor, used to execute the computer program to implement the steps of the wavefront recovery inference method for strong flicker perception without wavefront sensors as described above.
[0015] Fourthly, embodiments of this application provide a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the wavefront recovery inference method for strong flicker perception without wavefront sensors as described above.
[0016] As can be seen from the above technical solutions, compared with the prior art, this application has the following advantages: This application provides a sensorless wavefront recovery inference method for strong scintillation sensing, applicable to wavefront recovery and compensation in free-space optical communication, adaptive optics reception, single-mode fiber coupling, structured light spatial mode transmission, or free-space quantum communication. First, the first target light intensity distribution on the first measurement plane at the receiver is obtained. Then, this first target light intensity distribution is input into a target wavefront prediction model to obtain the corresponding estimated target phase. The target wavefront prediction model is trained offline using a dual-plane light intensity self-supervised method. This method utilizes two sets of measured light intensity distributions at different transmission distances to construct a light field propagation consistency loss, iteratively optimizing model parameters without requiring an external wavefront sensor for phase labeling. Finally, a conjugate correction phase is determined based on the estimated target phase, and the target light wave is corrected using this phase. Thus, by training the wavefront prediction model using a dual-plane light intensity self-supervised method, wavefront recovery and conjugate correction are achieved without an external wavefront sensor, utilizing only a single-plane light intensity distribution. This adapts to atmospheric turbulence and strong scintillation scenarios, improving the receiving coupling efficiency of free-space optical communication. Attached Figure Description
[0017] Figure 1 A flowchart of a wavefront recovery inference method for strong flicker sensing without a wavefront sensor provided in an embodiment of this application; Figure 2 A schematic diagram of an adaptive optics receiving hardware optical path and a model self-supervised training information flow provided in an embodiment of this application; Figure 3 This is a schematic diagram of a wavefront recovery inference system for strong flicker sensing without a wavefront sensor, provided in an embodiment of this application. Detailed Implementation
[0018] As mentioned earlier, existing technologies lead to a decrease in the coupling efficiency of free-space optical communication receivers. Specifically, current wavefront correction methods generally rely on wavefront sensors to collect phase information for wavefront correction. However, wavefront sensors are prone to detection failure under strong scintillation conditions. Furthermore, some wavefront-sensorless solutions depend on true phase value labeling or pure simulation-supervised training, lacking constraints from measured optical field propagation. This results in limited wavefront inversion accuracy and difficulty in effectively implementing conjugate correction, ultimately leading to a decrease in the coupling efficiency of free-space optical communication receivers.
[0019] To address the aforementioned issues, this application provides a wavefront recovery inference method for strong flicker sensing without a wavefront sensor, comprising: firstly, acquiring the first target light intensity distribution of the target light wave on a first measurement plane at the receiving end; then, inputting the first target light intensity distribution into a target wavefront prediction model to obtain the corresponding estimated target phase; wherein, the target wavefront prediction model is obtained through offline training using a dual-plane light intensity self-supervised method; the dual-plane light intensity self-supervised method utilizes two sets of measured light intensity distributions at different transmission distances to construct a light field propagation consistency loss, iteratively optimizing model parameters, and requires no external wavefront sensor to provide phase annotation; finally, determining a conjugate correction phase based on the estimated target phase, and using the conjugate correction phase to correct the target light wave.
[0020] Thus, by training the wavefront prediction model through a dual-plane light intensity self-supervised method, wavefront recovery and conjugate correction can be achieved using only a single-plane light intensity distribution without the need for external wavefront sensors. This adapts to atmospheric turbulence and strong scintillation scenarios and improves the receiving coupling efficiency of free-space optical communication.
[0021] It should be noted that the wavefront recovery inference method and related products for strong flicker sensing without a wavefront sensor provided in this application embodiment can be applied to the field of free-space optical communication technology. The above is merely an example and does not limit the application field of the wavefront recovery inference method and related products for strong flicker sensing without a wavefront sensor provided in this application.
[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0023] Figure 1 This is a flowchart illustrating a wavefront recovery inference method for strong flicker sensing without a wavefront sensor, provided in an embodiment of this application. (Combined with...) Figure 1 As shown in the embodiments of this application, a wavefront recovery inference method for strong flicker sensing without a wavefront sensor may include: S101: Obtain the first target light intensity distribution of the target light wave on the first measurement plane of the receiver.
[0024] In practical applications, it is first necessary to obtain the first target light intensity distribution of the target light wave (the distorted light wave to be corrected) on the first measurement plane at the receiving end. The first measurement plane is either the pupil plane or a plane conjugate to the pupil plane. The pupil plane is the plane corresponding to the aperture stop in the optical system that restricts the propagation of the light beam. The plane conjugate to the pupil plane is the observation plane where the wavefront distribution after optical imaging corresponds one-to-one with the pupil plane. The first measurement plane acquires light intensity information from the first camera, specifically the light intensity information of the target light wave transmitted to the first measurement plane after atmospheric turbulence disturbance, and obtains the corresponding two-dimensional planar light intensity distribution data (first target light intensity distribution). This step only requires single-plane light intensity acquisition to achieve subsequent wavefront recovery, eliminating the need for additional wavefront sensors to collect phase information, thus reducing hardware deployment costs and system complexity. Simultaneously, this approach avoids phase measurement distortion caused by wavefront sensor failure under strong scintillation conditions, providing reliable observational data for subsequent wavefront prediction models and supporting the subsequent wavefront recovery and optical wave correction processes.
[0025] S102: Input the first target light intensity distribution into the target wavefront prediction model and obtain the corresponding target estimated phase; the target wavefront prediction model is obtained through offline training by a dual-plane light intensity self-supervised method; the dual-plane light intensity self-supervised method uses two sets of plane measured light intensity distributions with different transmission distances to construct the light field propagation consistency loss, iteratively optimizes the model parameters, and does not require an external wavefront sensor to provide phase labeling.
[0026] Furthermore, the target wavefront prediction model is ZeMS-Net; The ZeMS-Net includes: a Swing-Transformer encoder, a Zernike prediction branch, a residual decoding branch, and a FiLM module; During offline training, a stopping gradient operation is performed on the conditional vector input to the FiLM module.
[0027] In practical applications, the light intensity distribution of the first target is obtained. Then, it is input into the target wavefront prediction model to obtain the corresponding estimated target phase. The target wavefront prediction model is trained offline using a dual-plane intensity self-supervised method. During training, two sets of measured light intensity distributions collected at different transmission distances are used to construct a light field propagation consistency loss. This loss is then used to iteratively optimize the model parameters, eliminating the need for external wavefront sensors to provide phase annotations. In this embodiment, the target wavefront prediction model uses ZeMS-Net, which includes a Swin-Transformer encoder for extracting multi-scale spatial features from the input intensity image, a Zernike prediction branch for estimating low-order Zernike mode coefficients, a residual decoding branch for estimating the local high-spatial-frequency residual phase, and a FiLM module. During offline training, a gradient-stopping operation is performed on the conditional vector input to the FiLM module to prevent the conditional vector from participating in gradient backpropagation, reducing the interference of high-frequency error signals from the residual decoding branch on the optimization of the Zernike prediction branch. Specifically, the input intensity image first enters a hierarchical Swin-Transformer encoder to capture global and anisotropic spatial correlation features caused by atmospheric turbulence. The Swin-Transformer encoder can adjust the number of layers, window size, and number of channels according to computational power and latency requirements. At the encoder bottleneck, the network is divided into a Zernike prediction branch and a residual decoding branch. The Zernike prediction branch mainly undertakes the task of predicting low-order physical aberrations and outputs low-order mode coefficients. And based on the low-order mode coefficients The formula for synthesizing a global reference wavefront is as follows: ; In the formula, As a global reference wavefront, This represents the total number of selected low-order modes. For Zernike mode order index, These are low-order modal coefficients. For the first Zernike basis functions of order 1. The residual decoding branch mainly undertakes the task of high-frequency local distortion recovery, outputting a high spatial frequency local residual wavefront. The final estimated phase output by the network is expressed as follows: ; In the formula, For the complete estimated phase of the network's final output, This is the high spatial frequency local residual wavefront output by the residual decoding branch. To guide low-order global aberration priors to high-frequency residual estimation, the FiLM module uses low-order modal coefficients... The conditional vector consisting of the maximum and minimum values of the global reference wavefront. Affine parameters related to the generation stage and And modulate the intermediate features in the residual decoding branch. The corresponding modulation expression is as follows: ; In the formula, These are intermediate features in the residual decoding branch. To scale the affine parameters, For offset affine parameters, This represents the characteristic-level linear modulation operation. This serves as the index subscript for the feature map channel. Thus, this scheme eliminates the need for ground truth phase annotation, reducing acquisition costs and avoiding annotation distortion caused by wavefront sensor failure under strong scintillation conditions. Simultaneously, ZeMS-Net combines optical priors and feature modulation mechanisms to improve wavefront reconstruction accuracy, while stopping gradient operations stabilizes conditional constraints, improving model training stability.
[0028] Furthermore, since the specific implementation methods of the dual-plane light intensity self-monitoring method are not entirely the same, the embodiments of this application can describe one possible implementation method.
[0029] In one case, the dual-plane light intensity self-monitoring method includes: The first training light intensity distribution of the training light wave is collected on the first measurement plane, and the second training light intensity distribution of the training light wave is collected simultaneously on the second measurement plane; the transmission distance between the first measurement plane and the second measurement plane is calibrated. The first training light intensity distribution is input into the initial wavefront prediction model to obtain the training estimated phase; A first estimated complex optical field is constructed by combining the first measured amplitude corresponding to the first training light intensity distribution and the training estimated phase, and the first estimated complex optical field is transmitted to the second measurement plane through the ASM model to obtain the second estimated complex optical field; The parameters of the initial wavefront prediction model are updated based on the photometric consistency loss between the estimated propagation amplitude corresponding to the second estimated complex optical field and the second measured amplitude corresponding to the second training light intensity distribution.
[0030] Furthermore, the photometric consistency loss is the L2 norm loss between the estimated propagation amplitude and the second measured amplitude.
[0031] In practical applications, during the training phase, the first camera first captures the training light wave (an arbitrary distorted beam input during the training phase) and its intensity distribution on the first measurement plane at the free-space optical link receiver. The second camera synchronously acquires the second training light intensity distribution of the second measurement plane, which is spaced at a calibration transmission distance (calibrated free space propagation distance z) from the first measurement plane. The first and second cameras can be two-dimensional light intensity detectors adapted to the target's operating wavelength. For example, when the wavelength of the training light wave is 1550 nm, the first and second cameras can be Hamamatsu indium gallium arsenide near-infrared cameras, model C14041-10U. Then, the intensity distribution of the first training light... Input the initial wavefront prediction model (untrained ZeMS-Net) and obtain the intensity distribution compared to the first training model. The corresponding training phase estimation This will then be used to estimate the phase during training. Compared with the first measured amplitude Combine the data to construct the first estimated complex optical field. And the first estimated complex optical field is obtained by using the ASM model. The light propagates to the second measurement plane, resulting in the second estimated complex optical field. Finally, based on the second estimated complex optical field... Corresponding estimated propagation amplitude The second measured amplitude corresponding to the second training light intensity distribution The photometric consistency loss between the two phases is used to update the parameters of the initial wavefront prediction model. Here, the photometric consistency loss is the estimated propagation amplitude. With the second measured amplitude The L2 norm loss between them is expressed as follows: ; In the formula, This refers to the loss of photometric uniformity (L2 norm loss). For the first training light intensity distribution, This is the first measured amplitude. This represents the true phase of the first measurement plane (used only for theoretical description and not labeled during training). For training purposes, the estimated phase, For the first estimated complex optical field, For the true complex optical field of the first measurement plane, For ASM operators, To ensure the true propagation of the operator, This indicates taking the modulus of the complex optical field after propagation. To calibrate the free space propagation distance (calibrate the transmission distance). Indicates L2 norm loss, It is a natural constant. This is the imaginary unit. Using the second measured amplitude Replacement. Understandably, this L2 norm loss simulates the first estimated complex optical field, composed of the first measured amplitude and the estimated phase used for training, through angular spectrum propagation to the second measurement plane. The simulated amplitude is compared with the second measured amplitude using a L2 norm squared comparison, utilizing the dual-plane light intensity propagation consistency constraint to achieve self-supervised training. Furthermore, calibrating the free-space propagation distance z requires simultaneously satisfying the discrete Rytov variance constraint and the spatial sampling anti-aliasing constraint. In this embodiment, z = 100 m can be used, and regardless of whether the sample set is generated by simulation or obtained from optical path measurement, it is uniformly divided into three categories: training samples, validation samples, and test samples, each used separately. Additionally, this training phase can employ a gradient optimization algorithm, using the Adaptive Moment Estimation (Adam) optimizer, setting the batch size to 16, and employing a learning rate strategy of linear warm-up followed by cosine annealing after 10 iteration cycles. The initial learning rate is... Minimum learning rate Thus, this scheme relies on the synchronous measured light intensity of two planes to construct self-supervised constraints, and can carry out training without the need for phase truth labeling provided by wavefront sensors. This can reduce hardware and data acquisition costs and avoid training bias caused by phase labeling distortion in strong scintillation environments. The angular spectrum propagation model can accurately simulate the light field diffraction propagation process, and the use of L2 loss can intuitively quantify amplitude differences, stably driving model iterative optimization to improve wavefront reconstruction accuracy.
[0032] In addition to the mainstream method of obtaining measured training samples through physical optical path acquisition, a numerical simulation scheme can also be used to generate paired dual-plane light intensity distributions in batches during the construction of training samples. These simulated samples can be used independently for model pre-training or mixed with measured samples to complete offline model training. This simulation scheme relies on the Split-Step Beam Propagation Method (SS-BPM) to simulate the complete atmospheric turbulence propagation process and generates a turbulent phase screen based on the Kolmogorov power spectrum to reproduce the random fluctuations in atmospheric refractive index. The entire simulation operation consists of three steps: first, performing a Fourier transform on the complex optical field; second, superimposing the turbulent phase screen to complete distortion modulation; and finally, performing an inverse Fourier transform to restore the spatial optical field. To ensure that the simulation results conform to real physical laws and avoid numerical calculation divergence, the discrete propagation step size must simultaneously satisfy the discrete Rytov variance constraint and the spatial sampling anti-aliasing constraint. As an example, an atmospheric refractive index structure constant can be set. The operating wavelength of the light wave is 1550 nm, and the corresponding calibration free-space propagation distance z is 100 m. The final simulation dataset contains a total of 20,000 pairs of dual-plane light intensity samples, which are then split in a 7:1.5:1.5 ratio, with 14,000 pairs as training samples, 3,000 pairs as validation samples, and 3,000 pairs as test samples. All the above numerical simulation calculations can be completed using general-purpose scientific computing software, such as the numerical computing library (NumPy) and the scientific computing library (SciPy), to perform wave optical diffraction derivation and spectral transformation, as well as all other light field simulation calculations.
[0033] Furthermore, since the methods of collecting training light intensity distributions are not entirely the same, this application embodiment can describe one possible collection method.
[0034] In one scenario, the acquisition of the first training light intensity distribution of the training light wave on the first measurement plane, and the simultaneous acquisition of the second training light intensity distribution of the training light wave on the second measurement plane, includes: The training light wave is split to obtain a first beam and a second beam; the first beam is transmitted along a first optical path to a first measurement plane, and the second beam is transmitted along a second optical path to a second measurement plane; The first training light intensity distribution corresponding to the training light wave is acquired in the first measurement plane; The second training light intensity distribution corresponding to the training light wave is acquired in the second measurement plane.
[0035] In practical applications, the training light wave is first split using a beam splitter to obtain a first beam and a second beam. The first beam is then transmitted along a first optical path to a first measurement plane, and the second beam is transmitted along a second optical path to a second measurement plane. Subsequently, a first camera acquires the intensity distribution of the first training light on the first measurement plane. The second training light intensity distribution is obtained by acquiring data from the second camera on the second measurement plane. This beam splitting acquisition method ensures that the two sets of light intensity data come from the same distorted light field, avoiding the differences in light field caused by different beams and different acquisition times. It provides synchronously matched observation data for constructing consistency constraints for dual-plane propagation, ensuring the self-supervised training effect of the model.
[0036] Figure 2 This is a schematic diagram of an adaptive optics receiving hardware optical path and a model self-supervised training information flow provided in an embodiment of this application. Combined with... Figure 2As shown, during the training phase, on the hardware side, distorted light waves can be input as training light waves through a turbulence chamber or simulated turbulence. The distorted light waves enter the first beam splitter and are split into two beams. One beam is directly sent to the spatial light modulator, and the corrected output beam enters the fiber coupling. The other beam enters the second beam splitter and is split into two beams again, which are transmitted to a first camera and a second camera with a calibrated free-space propagation distance, forming a dual-plane intensity measurement module. The light intensity images acquired by the first and second cameras are sent to the correction network (ZeMS-Net). The network calculates the corrected phase and then sends it back to the spatial light modulator as an electrical signal to complete the wavefront correction. On the information flow side, there are two parallel links: physical flow and numerical flow. The physical flow is the light field corresponding to the input distorted light wave. →First training light intensity distribution measured by the first camera (True phase of the first measurement plane) (Unknown), the light beam continues to physically propagate a distance → the second training light intensity distribution measured by the second camera. Numerical flow is used to apply the measured first training light intensity distribution. Feed into the calibration network (ZeMS-Net) → Output estimated phase for training Using the first measured amplitude and training phase estimation Construct the first estimated complex optical field, then use the ASM model to propagate the complex optical field forward by the same distance in the computer to obtain the predicted second estimated complex optical field. Finally, based on the second estimated complex optical field Corresponding estimated propagation amplitude With the second training light intensity distribution The corresponding second measured amplitude The loss function is calculated, and then the parameters of ZeMS-Net are iteratively updated based on the backpropagation of the loss gradient.
[0037] S103: Determine the conjugate correction phase based on the target estimated phase, and use the conjugate correction phase to correct the target light wave.
[0038] In practical applications, the target estimated phase is obtained. Then, the phase is estimated based on this target. Determine the corresponding conjugate correction phase - Then use conjugate correction phase - Phase compensation correction is performed on the target light wave. In this way, distortion compensation is completed by constructing a conjugate correction phase from the target estimated phase output by the model. This enables sensorless adaptive optics correction based on the aforementioned single-plane light intensity acquisition and wavefront prediction scheme, eliminating the hardware dependence on wavefront sensors in traditional schemes, while avoiding the problem of correction performance degradation caused by phase detection failure under strong scintillation conditions.
[0039] Furthermore, since the methods of optical wave correction are not entirely the same, the embodiments of this application can be used to describe one possible correction method.
[0040] In one instance, correcting the target light wave using the conjugate correction phase includes: The conjugate correction phase is applied to a spatial light modulator or deformable mirror to correct the target light wave.
[0041] In practical applications, the obtained conjugate correction phase can be... By loading the light onto a spatial light modulator or deformable mirror, and leveraging the modulator's or mirror's ability to manipulate the light wave phase in real time, phase compensation correction can be applied to the target light wave. Furthermore, a third camera or coupling detection module can be added to record the corrected beam. Thus, using a spatial light modulator or deformable mirror as the phase correction actuator allows for flexible adaptation to different adaptive optics system hardware architectures, enabling rapid response to the conjugate correction phase output from the model, real-time compensation for wavefront distortion introduced by atmospheric turbulence, and improved beam quality of the target light wave.
[0042] Furthermore, to fully verify the wavefront correction performance of the ZeMS-Net in this application, the single-mode fiber coupling efficiency is used as the core evaluation index. This index is calculated by the overlap integral of the corrected complex optical field and the fundamental mode field of the single-mode fiber, which can objectively quantify the wavefront recovery accuracy and beam coupling improvement effect. Multiple turbulence simulations and physical experiments show that at a transmission distance of 2000m and with the atmospheric refractive index structure constant at a certain level... to In typical turbulent scenarios, the average fiber coupling efficiency of the uncorrected beam is only 16.44%, while the coupling efficiency after correction using ZeMS-Net reaches 47.99%, significantly better than the 33.93% of the traditional Gerchberg-Saxton (GS) iterative phase retrieval algorithm and the 39.11% of the general convolutional baseline neural network, demonstrating superior wavefront reconstruction capabilities in typical turbulent conditions. This is further demonstrated at a transmission distance of 4000m and a Ritov variance... Under extreme conditions of strong scintillation, traditional self-supervised models constrained by ideal Gaussian amplitudes essentially lose their correction capabilities, with coupling efficiency dropping to 29.12%, roughly the same as the uncorrected condition (29.13%). However, the ZeMS-Net scheme used in this application, relying on measured amplitude injection and multi-branch feature modulation mechanisms, maintains a high coupling correction performance of 55.86%, effectively verifying the crucial supporting role of measured amplitude priors in strong scintillation distortion correction. Physical platform test results further demonstrate that this scheme can improve the beam Strell ratio from 0.38 in the uncorrected state to 0.70, reduce the root mean square error of light intensity from 1.38 to 0.66, and significantly improve beam quality. Simultaneously, the model's single inference latency is only 2.90 ± 0.30 ms, and the computation frame rate reaches 344.5 FPS, demonstrating fast real-time wavefront correction capabilities. It should be noted that the above experimental results only represent the performance of this application under the corresponding implementation conditions and do not constitute a performance limitation for all application scenarios of this application.
[0043] In summary, this application first obtains the first target light intensity distribution on the first measurement plane of the target light wave at the receiving end. Then, the first target light intensity distribution is input into the target wavefront prediction model to obtain the corresponding estimated target phase. The target wavefront prediction model is trained offline using a dual-plane light intensity self-supervised method. This method utilizes two sets of measured light intensity distributions at different transmission distances to construct a light field propagation consistency loss, iteratively optimizing model parameters, and requires no external wavefront sensor to provide phase annotation. Finally, a conjugate correction phase is determined based on the estimated target phase, and the target light wave is corrected using this conjugate correction phase. Thus, by training the wavefront prediction model using a dual-plane light intensity self-supervised method, wavefront recovery and conjugate correction are achieved without an external wavefront sensor, utilizing only a single-plane light intensity distribution. This adapts to atmospheric turbulence and strong scintillation scenarios, improving the receiving coupling efficiency of free-space optical communication.
[0044] Figure 3 This is a schematic diagram of a wavefront recovery inference system for strong flicker sensing without a wavefront sensor, provided in an embodiment of this application. (Combined with...) Figure 3 As shown, the wavefront recovery inference system 300 without a wavefront sensor includes: The acquisition module 301 is used to acquire the first target light intensity distribution of the target light wave on the first measurement plane of the receiving end; The prediction module 302 is used to input the first target light intensity distribution into the target wavefront prediction model and obtain the corresponding target estimated phase; the target wavefront prediction model is obtained through offline training by a dual-plane light intensity self-supervised method; the dual-plane light intensity self-supervised method uses two sets of plane measured light intensity distributions with different transmission distances to construct the light field propagation consistency loss, iteratively optimizes the model parameters, and does not require an external wavefront sensor to provide phase labeling; The correction module 303 is used to determine a conjugate correction phase based on the target estimated phase, and to use the conjugate correction phase to correct the target light wave.
[0045] The target wavefront prediction model is ZeMS-Net. The ZeMS-Net includes: a Swing-Transformer encoder, a Zernike prediction branch, a residual decoding branch, and a FiLM module; During offline training, a stopping gradient operation is performed on the conditional vector input to the FiLM module.
[0046] As one implementation method, regarding how to achieve the dual-plane light intensity self-supervision approach, the aforementioned wavefront recovery inference system 300 without a wavefront sensor further includes: a training module; the training module includes: The acquisition unit is used to acquire the first training light intensity distribution of the training light wave on the first measurement plane, and simultaneously acquire the second training light intensity distribution of the training light wave on the second measurement plane; the first measurement plane and the second measurement plane are calibrated with a transmission distance between them; The input unit is used to input the first training light intensity distribution into the initial wavefront prediction model and obtain the training estimated phase. The estimation unit is used to construct a first estimated complex optical field by combining the first measured amplitude corresponding to the first training light intensity distribution and the training estimated phase, and to transmit the first estimated complex optical field to the second measurement plane through the ASM model to obtain a second estimated complex optical field. The update unit is used to update the parameters of the initial wavefront prediction model based on the photometric consistency loss between the estimated propagation amplitude corresponding to the second estimated complex optical field and the second measured amplitude corresponding to the second training light intensity distribution.
[0047] Furthermore, the photometric consistency loss is the L2 norm loss between the estimated propagation amplitude and the second measured amplitude.
[0048] As one implementation method, regarding how to acquire the first training light intensity distribution and the second training light intensity distribution, the aforementioned acquisition unit is specifically used for: The training light wave is split to obtain a first beam and a second beam; the first beam is transmitted along a first optical path to a first measurement plane, and the second beam is transmitted along a second optical path to a second measurement plane; The first training light intensity distribution corresponding to the training light wave is acquired in the first measurement plane; The second training light intensity distribution corresponding to the training light wave is acquired in the second measurement plane.
[0049] As one implementation method, regarding how to correct the target light wave, the aforementioned correction module 303 is specifically used for: The conjugate correction phase is applied to a spatial light modulator or deformable mirror to correct the target light wave.
[0050] In summary, this application first obtains the first target light intensity distribution on the first measurement plane of the target light wave at the receiving end. Then, the first target light intensity distribution is input into the target wavefront prediction model to obtain the corresponding estimated target phase. The target wavefront prediction model is trained offline using a dual-plane light intensity self-supervised method. This method utilizes two sets of measured light intensity distributions at different transmission distances to construct a light field propagation consistency loss, iteratively optimizing model parameters, and requires no external wavefront sensor to provide phase annotation. Finally, a conjugate correction phase is determined based on the estimated target phase, and the target light wave is corrected using this conjugate correction phase. Thus, by training the wavefront prediction model using a dual-plane light intensity self-supervised method, wavefront recovery and conjugate correction are achieved without an external wavefront sensor, utilizing only a single-plane light intensity distribution. This adapts to atmospheric turbulence and strong scintillation scenarios, improving the receiving coupling efficiency of free-space optical communication.
[0051] In addition, this application also provides a wavefront recovery inference device for strong flicker perception without a wavefront sensor, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the wavefront recovery inference method for strong flicker perception without a wavefront sensor as described above.
[0052] In addition, this application also provides a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the wavefront recovery inference method for strong flicker perception without wavefront sensors as described above.
[0053] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A strong-flare-aware wavefront sensor wavefront-recovery inference method, comprising: The method includes: Acquire the first target light intensity distribution of the target light wave on the first measurement plane of the receiver; The first target light intensity distribution is input into the target wavefront prediction model to obtain the corresponding target estimated phase; the target wavefront prediction model is trained offline through a dual-plane light intensity self-supervised method; the dual-plane light intensity self-supervised method uses two sets of plane measured light intensity distributions with different transmission distances to construct the light field propagation consistency loss, iteratively optimizes the model parameters, and does not require an external wavefront sensor to provide phase labeling; The conjugate correction phase is determined based on the target estimated phase, and the target light wave is corrected using the conjugate correction phase.
2. The method of claim 1, wherein, The dual-plane light intensity self-monitoring method includes: The first training light intensity distribution of the training light wave is collected on the first measurement plane, and the second training light intensity distribution of the training light wave is collected simultaneously on the second measurement plane; the transmission distance between the first measurement plane and the second measurement plane is calibrated. The first training light intensity distribution is input into the initial wavefront prediction model to obtain the training estimated phase; A first estimated complex optical field is constructed by combining the first measured amplitude corresponding to the first training light intensity distribution and the training estimated phase, and the first estimated complex optical field is transmitted to the second measurement plane through the ASM model to obtain the second estimated complex optical field; The parameters of the initial wavefront prediction model are updated based on the photometric consistency loss between the estimated propagation amplitude corresponding to the second estimated complex optical field and the second measured amplitude corresponding to the second training light intensity distribution.
3. The method of claim 2, wherein, The photometric consistency loss is the L2 norm loss between the estimated propagation amplitude and the second measured amplitude.
4. The method according to claim 1, characterized in that, The target wavefront prediction model is ZeMS-Net.
5. The method according to claim 4, characterized in that, The ZeMS-Net includes: a Swing-Transformer encoder, a Zernike prediction branch, a residual decoding branch, and a FiLM module; During offline training, a stopping gradient operation is performed on the conditional vector input to the FiLM module.
6. The method according to claim 2, characterized in that, The acquisition of the first training light intensity distribution of the training light wave on the first measurement plane, and the simultaneous acquisition of the second training light intensity distribution of the training light wave on the second measurement plane, includes: The training light wave is split to obtain a first beam and a second beam; the first beam is transmitted along a first optical path to a first measurement plane, and the second beam is transmitted along a second optical path to a second measurement plane; The first training light intensity distribution corresponding to the training light wave is acquired in the first measurement plane; The second training light intensity distribution corresponding to the training light wave is acquired in the second measurement plane.
7. The method according to claim 1, characterized in that, The step of correcting the target light wave using the conjugate correction phase includes: The conjugate correction phase is applied to a spatial light modulator or deformable mirror to correct the target light wave.
8. A wavefront recovery inference system without wavefront sensor for strong flicker sensing, characterized in that, include: The acquisition module is used to acquire the first target light intensity distribution of the target light wave on the first measurement plane of the receiving end; The prediction module is used to input the first target light intensity distribution into the target wavefront prediction model and obtain the corresponding target estimated phase; the target wavefront prediction model is trained offline through a dual-plane light intensity self-supervised method. The dual-plane light intensity self-supervised method uses two sets of measured light intensity distributions on two planes with different transmission distances to construct the light field propagation consistency loss, iteratively optimizes the model parameters, and does not require external wavefront sensors to provide phase labeling. The correction module is used to determine a conjugate correction phase based on the target estimated phase, and to correct the target light wave using the conjugate correction phase.
9. A wavefront recovery inference device without a wavefront sensor for strong flicker sensing, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the wavefront recovery inference method for strong flicker sensing without wavefront sensors as described in any one of claims 1 to 7 when executing the computer program.
10. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the steps of the wavefront recovery inference method for strong flicker sensing without wavefront sensors as described in any one of claims 1 to 7.