A method for controlling a wavefront probeless adaptive optical system

Abstract

The present application relates to the technical field of adaptive optics, and discloses a wavefront-detection-free adaptive optical system control method, which uses a camera to collect an extended target light intensity distribution image transmitted through atmospheric turbulence, inputs the image into a deep learning target detection model, extracts a target recognition confidence as a performance index of closed-loop control, and controls a controller to estimate a performance index gradient by applying bilateral random voltage disturbance based on a stochastic parallel gradient descent algorithm, and update control voltage of a corrector to drive an actuator to compensate for wavefront distortion according to the performance index gradient. The present application discloses a direct mapping between optical imaging quality and machine vision recognition rate by using deep learning confidence, effectively solves the poor applicability of traditional evaluation indexes in the extended target scene, and significantly improves the imaging clarity and recognition accuracy of the system in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of adaptive optics technology, specifically to a control method for an adaptive optics system without wavefront detection. Background Technology

[0002] Adaptive optics technology effectively improves the imaging resolution of optical systems by compensating for atmospheric turbulence or inherent aberrations in real time, and is widely used in astronomical observation, space target surveillance, and laser communication. Traditional adaptive optics systems typically employ Hartmann wavefront sensors to measure wavefront distortion and use deformable mirrors for phase compensation. This closed-loop control architecture based on wavefront sensors performs exceptionally well when imaging point targets such as stars.

[0003] However, traditional adaptive optics technology faces limitations in imaging applications targeting extended targets such as drones and satellites. On one hand, wavefront sensors have strict requirements regarding the shape and brightness of the target being detected, relying on point beacons or high-contrast reference sources for operation. When the observed object is an extended target with uneven surface brightness distribution, or in environments with strong turbulence and low signal-to-noise ratios, wavefront sensors often fail to accurately measure wavefront slope information, leading to system instability or even malfunction. This hardware limitation increases system complexity and cost, and also restricts its application capabilities in complex scenarios.

[0004] To reduce reliance on wavefront sensors, wavefront-free adaptive optics, optimized based on performance metrics, has emerged. This technique typically uses image sharpness function, Strel ratio, or ring energy as evaluation metrics, iteratively optimizing the corrector surface shape through algorithms such as stochastic parallel gradient descent. While these traditional metrics, based on the statistical characteristics of image grayscale, have some effect on improving human visual perception, they only focus on low-level statistical features and fail to encompass high-level semantic information. In automated tasks for machine vision, improved image sharpness does not necessarily equate to increased target feature recognizability. This mismatch between evaluation metrics and the requirements of backend recognition tasks often results in corrected images that fail to meet the practical needs of high-precision target recognition and tracking. Therefore, establishing an evaluation mechanism that directly reflects the quality of machine vision recognition and thereby achieves closed-loop correction of the optical system is a pressing technical problem. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a wavefront-free adaptive optics system control method, aiming to solve the problems that existing adaptive optics systems rely on wavefront sensors when imaging extended targets, resulting in poor environmental adaptability, and that traditional image evaluation metrics cannot accurately reflect machine vision recognition quality, making it difficult for the corrected images to meet the requirements of automatic target recognition and tracking.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a control method for a wavefront-free adaptive optics system. The system employing this method includes an extended target, a combined optical path beam-shrinking system, a corrector, a high-voltage amplifier driver, an imaging lens, a camera, and a controller. The control method includes the following steps: The camera acquires an image of the light intensity distribution of the extended target after reflection by the combined optical path beam-shrinking system and the corrector. The light intensity distribution image contains wavefront distortion information after transmission through atmospheric turbulence and is transmitted to the controller. The controller inputs the received light intensity distribution image into a preset deep learning target detection model, which extracts features and locates the extended target, and outputs the detection result containing the target detection box and probability information. The controller extracts the detection box information with the highest confidence from the detection results, calculates the target recognition confidence, and sets the target recognition confidence as the performance index of the adaptive optics closed-loop control. The controller executes a stochastic parallel gradient descent algorithm, applies a bilateral random perturbation voltage to the corrector, and estimates the gradient direction of the performance index based on the performance index change value corresponding to the image acquired by the camera under the perturbation. The controller calculates the updated control voltage vector based on the estimated gradient direction, and drives the actuator of the corrector to change the mirror surface shape through the high-voltage amplifier driver to compensate for wavefront distortion until the performance index reaches a convergent state.

[0007] Preferably, during the acquisition of the light intensity distribution image, the static aberration environment corresponding to the wavefront distortion information is constructed by a Gaussian random phase perturbation model. The Gaussian random phase perturbation model is generated based on the Gaussian blur kernel function and uses the standard deviation parameter to define the degree of blur of the image.

[0008] Preferably, during the acquisition of the light intensity distribution image, the dynamic aberration environment corresponding to the wavefront distortion information is constructed through the Ornstein-Uhlenbeck process, wherein the evolution of the dynamic blur intensity follows a stochastic differential equation, which is determined by the regression rate, the expected blur mean, and the noise intensity coefficient, and is used to simulate the mean regression characteristics and stochastic fluctuation characteristics of atmospheric turbulence intensity.

[0009] Preferably, the deep learning target detection model includes a backbone network, a neck component, and a head component connected in sequence. The backbone network adopts a convolutional neural network structure to extract multi-scale feature maps from the light intensity distribution image; the neck component adopts a feature fusion network structure to aggregate the deep semantic information and shallow detail information output by the backbone network; the head component contains multiple detection heads to generate detection results based on the aggregated feature maps, and the data structure of the detection results includes bounding box coordinates, target presence probability, and target category probability.

[0010] Preferably, the performance index is calculated as the product of the target presence probability and the target category probability. The target presence probability represents the model's confidence that the current detection box contains an object of interest, while the target category probability represents the conditional probability that the object belongs to a preset category given the confirmed presence of the object.

[0011] Preferably, the step of estimating the gradient direction of the performance index specifically includes: generating a set of random voltage perturbation vectors that satisfy a Bernoulli distribution, the dimension of which is consistent with the number of driving units of the corrector; adding the current base voltage vector to the random voltage perturbation vector to obtain a positive perturbation voltage, applying it to the corrector, acquiring a positive perturbation image and calculating a positive performance index; subtracting the base voltage vector from the random voltage perturbation vector to obtain a negative perturbation voltage, applying it to the corrector, acquiring a negative perturbation image and calculating a negative performance index; calculating the difference between the positive and negative performance indices, and using this difference as the basis for characterizing the gradient information of the performance index in the current voltage space.

[0012] Preferably, the update rule for the corrector control voltage vector is as follows: the base voltage vector is added to the correction term to obtain the updated control voltage vector; the correction term is determined by the product of the learning rate, the aforementioned difference, and the random voltage disturbance vector, wherein the learning rate is used as the step size for control voltage update.

[0013] Preferably, when calculating the target recognition confidence score, if the deep learning target detection model does not contain any target detection box in the detection result at the current moment, the controller sets the target recognition confidence score to zero to correspond to the state where the target features are lost due to severe image distortion.

[0014] Preferably, the deep learning object detection model is trained using a transfer learning strategy, the weights are initialized using a pre-trained weight file, the training dataset is a public dataset containing drone images, and the class labels during training are limited to a single class.

[0015] Preferably, the controller is connected to the high-voltage amplifier driver via a D / A conversion interface. The high-voltage amplifier driver receives the analog voltage signal from the controller and amplifies it. The amplified voltage signal is then transmitted to each actuator electrode of the corrector to drive the mirror to deform and modulate the beam phase.

[0016] This invention provides a control method for a wavefront-free adaptive optics system. It offers the following advantages: 1. This invention uses the target recognition confidence score output by a deep learning target detection model as the closed-loop feedback performance index of the adaptive optics system. This index is jointly determined by the target presence probability and the target category probability. Compared with traditional optical evaluation functions such as Strel ratio or ring energy, this scheme establishes a direct mapping relationship between wavefront correction and machine vision recognition tasks, effectively solving the problem of poor correlation between traditional indices and target recognition rate in extended target imaging. This makes the corrected image features more conducive to extraction and classification by computer algorithms.

[0017] 2. This invention utilizes light intensity distribution images acquired by a camera in conjunction with a stochastic parallel gradient descent algorithm for wavefront-free detection control, eliminating the need for traditional direct phase detection components such as Hartmann wavefront sensors. This architecture avoids the situation where wavefront sensor detection accuracy decreases or even fails under extended target imaging or strong turbulence conditions due to uneven target brightness distribution or low signal-to-noise ratio. It reduces the hardware complexity and optical path debugging difficulty of the system, and improves the system's stability in complex observation environments.

[0018] 3. This invention utilizes a stochastic parallel gradient descent algorithm to apply bilateral random perturbations to the corrector and update the control voltage, iteratively searching for the extreme points of the performance indicators. This control strategy can adjust the wavefront phase of the optical system in real time based on feedback from the deep learning model, continuously optimizing imaging sharpness in dynamically changing turbulent environments, thereby improving the system's real-time tracking capability and recognition accuracy for extended targets. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall structure of the wavefront-free adaptive optics iterative control algorithm with target recognition confidence as the performance index proposed in this invention; Figure 2 This is a diagram of the internal network structure of the YOLOv11 model; Figure 3 It is the image after blurring; Figure 4 This is the corrected image; Figure 5 It is a graph showing the change in target recognition confidence with the number of iterations under static aberrations; Figure 6These are five randomly selected images with dynamic aberration degradation. Figure 7 It is a curve showing how the confidence level of target recognition changes with the number of iterations under dynamic aberrations; Figure 8 It is a graph showing how variance and confidence level change with the number of iterations. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See attached document Figure 1 The present invention provides a wavefront-free adaptive optics system, which includes: an extended target, an atmospheric turbulence simulation environment, a combined optical path beam-shrinking system, a corrector, a high-voltage amplifier driver (integrated in the front end of the corrector), an imaging lens, a camera, and a controller.

[0022] exist Figure 1 The quadcopter is the extended target, which in this embodiment is a quadcopter drone, serving as the imaging object; Atmosphere turbulence is atmospheric turbulence; L1 and L2 are lens groups, which together constitute a combined optical path beam-shrinking system to reduce the beam aperture to match the corrector size; the corrector is a corrector, i.e., a deformable mirror, used to modulate the phase of the beam; L3 is an imaging lens, used to focus the corrected beam onto the camera target surface; the camera is a camera used to acquire light intensity distribution images; and the controller is a controller used to receive images and run control algorithms, outputting voltage to drive the corrector.

[0023] The corrector is used to modulate the wavefront phase of a light beam. In this embodiment, a 61-element deformable mirror is selected as the actuating device for the corrector. However, those skilled in the art will understand that the corrector can also be a liquid crystal spatial light modulator, a microelectromechanical system deformable mirror, or other wavefront correction devices capable of changing the phase of light waves, and the present invention does not limit this.

[0024] The light reflected or emitted by the extended target passes through an atmospheric turbulence simulation environment, causing distortion of the light wavefront. The distorted beam enters a combined optical path beam-shrinking system for beam aperture matching. The combined optical path beam-shrinking system contains two optical path beam-shrinking structures used to adjust the beam diameter to fit the aperture of subsequent optical components.

[0025] The beam, after being reduced in size, is projected onto the surface of the corrector. In this embodiment, a 61-element deformable mirror is used as the corrector. After being reflected and phase-modulated by the corrector, the beam is converged by the imaging lens. The converged beam is imaged onto the photosensitive surface of the camera. The camera is connected to the controller via an image acquisition card, transmitting the acquired digital image data to the controller in real time.

[0026] The controller is internally configured with a runtime environment based on Python 3.10 and the Ultralytics YOLO framework. The controller loads the pre-trained weight file YOLOv11n.pt. This weight file is generated based on the Drone YOLO Detection Dataset, using a single training class label "quadcopter" after transfer learning. The controller utilizes this environment to perform deep learning inference tasks.

[0027] This invention also provides a control method for an adaptive optics system without wavefront detection, which executes closed-loop iterative logic through a controller. The method first executes step S1, acquiring a distorted image of the extended target after atmospheric turbulence-simulated environmental disturbance via a camera. This distorted image reflects the imaging quality and wavefront distortion state of the optical system at the current moment.

[0028] Then, in step S2, the controller inputs the acquired distorted image into the deep learning neural network object detection model. The model performs feature extraction and inference on the image and outputs the object recognition confidence score.

[0029] Next, step S3 is executed, where the controller calculates the system performance indicators. This performance metric Based on the probability of the target's existence With target category probability The product is determined by the following formula: ; in, This indicates the probability that a target exists within the detection bounding box. This indicates the probability that the target belongs to a specific category (such as a drone). It serves as a basis for quantifying the current image imaging quality and is used to determine whether the iterative process has reached a convergence state.

[0030] Then, step S4 is executed, where the controller bases its calculations on the performance metrics. A wavefront-less adaptive optics iterative control algorithm is used to adjust the drive quantity of the corrector. In this embodiment, the algorithm employs stochastic parallel gradient descent (SPGD). The controller adjusts the drive quantity based on performance indicators. The gradient direction is estimated based on the change in the voltage vector, and the control voltage vector is updated according to the following rules: ; in, For the first The control voltage for the next iteration. For the first The control voltage for the next iteration. For learning rate, This is the voltage disturbance vector. This represents the difference in performance indicators under positive and negative disturbances, i.e. In this embodiment, the learning rate Set to 0.5, voltage disturbance vector Set to 0.5.

[0031] Finally, step S5 is executed, and the controller updates the control voltage. An actuation force is applied to the corrector, physically altering the optical path phase. The system repeats steps S3 to S5, continuously updating the image and calculating confidence levels until the performance metrics are met. Stable convergence enables sharpening correction of extended target imaging.

[0032] Based on the aforementioned system architecture, the method in this embodiment executes step S1, namely, distorted image acquisition and atmospheric turbulence simulation environment construction, through the controller coordinating with various hardware modules. This step aims to acquire a target image containing wavefront distortion information, providing raw data support for subsequent closed-loop correction.

[0033] In terms of image acquisition mechanism, the camera is at the end of the optical imaging link, responsible for receiving the beam after beam matching by the combined optical path beam-shrinking system and phase modulation by the corrector. Unlike traditional adaptive optics systems that rely on wavefront sensors (such as Shaker-Hartmann sensors) to directly measure wavefront slope or phase information, this embodiment completely abandons the wavefront detection branch. The system uses the light intensity distribution image acquired by the camera as the sole input feedback source for the closed-loop control loop. The controller reads the digital image signal output by the camera through the acquisition card. The image object acquired at this time is an extended target, that is, a non-point light source object with resolvable geometric shape and surface texture details. In this embodiment, the extended target is specifically represented by a drone. Since the light energy of the extended target is distributed in a certain area rather than concentrated at a point, and the surface brightness distribution is uneven, direct correction using image grayscale information is more complex than that for a point light source. Therefore, the acquired high-fidelity distorted image is the basis for subsequent confidence calculation.

[0034] To verify the system's adaptability to different atmospheric conditions, an atmospheric turbulence simulation environment needs to be constructed during the experimental or simulation phase. (See attached...) Figure 3 , Figure 3The image with added blur shows the imaging state of the extended target (UAV) after being subjected to simulated atmospheric turbulence disturbance. The image shows obvious blurring and loss of detail. This embodiment first uses a Gaussian random phase perturbation model to simulate the static aberration environment.

[0035] Specifically, the controller performs mathematical operations in the simulation environment to degrade the original sharp target image, simulating wavefront phase distortion that occurs when light waves pass through a static atmospheric turbulence layer. This simulation process employs a Gaussian-based blur algorithm, applying a point spread function to the original image through convolution operations. This Gaussian blurring process is executed according to the following formula: ; In the above formula, Represents the Gaussian blur kernel function in two-dimensional space coordinates The weight value at the location; The coordinates are the spatial coordinates with the center of the convolution kernel as the origin; Pi; Standard deviation is used to quantify the diffusion of a Gaussian distribution, i.e., a physical quantity that controls the degree of image blur.

[0036] Regarding parameter settings, The magnitude of the value is directly related to the intensity of atmospheric turbulence; A larger value indicates a flatter Gaussian distribution, a stronger frequency domain low-pass filtering effect, more severe loss of high-frequency details in the image, and a higher simulated turbulence intensity. In the static aberration experiment of this embodiment, the following values ​​were selected: This value is set as a baseline. The choice of this value is not arbitrary, but rather to simulate conditions of strong turbulence (corresponding to atmospheric coherence length). (In smaller cases) Severe image degradation scenarios. Under these parameters, the contour edges of the extended target become significantly blurred, and internal texture features are difficult to discern, thus constructing a highly challenging initial correction state to fully test the recognition limits of the subsequent YOLOv11 model in low-quality images and the convergence ability of the SPGD algorithm.

[0037] Through this numerical modeling, the present invention transforms the abstract atmospheric turbulence effect into a controllable and reproducible mathematical variable, ensuring the scientific rigor of the verification process. The specific implementation of image convolution processing and the calculation of the Gaussian function are standard techniques in the field of digital image processing and will not be detailed here.

[0038] See attached document Figure 6In real-world applications, atmospheric turbulence is not static but rather a dynamic, stochastic process that evolves rapidly over time. To verify the system's ability to track and correct time-varying aberrations, and especially its adaptability to extended targets in a dynamic, blurred environment, a stochastic process was directly used. The blur intensity value was varied over time, and by introducing a time-dimensional variable, the continuous change in light wave phase distortion caused by fluctuations in the refractive index of the real atmosphere was simulated.

[0039] The controller utilizes the Ornstein-Uhlenbeck (OU) process to generate a time-dependent dynamic fuzzy parameter sequence. The OU process is a stationary Gaussian-Markov process with mean-regression characteristics, capable of accurately describing the physical phenomenon of atmospheric turbulence intensity fluctuating around a certain mean. Dynamic fuzzy intensity The evolution follows a stochastic differential equation, specifically as follows: ; In the above formula, Indicates the next moment The fuzziness intensity value, Indicates the current time The fuzzy intensity value, the variation of which simulates the time-varying characteristics of turbulence intensity. The time step determines the time resolution of the simulation.

[0040] The parameters in the equations have explicit physical meanings: The regression rate quantifies the speed at which turbulence intensity recovers to equilibrium after deviating from the mean, reflecting the time-dependent length of atmospheric disturbance. The desired fuzzy mean, i.e., the long-term average level of turbulence intensity, is usually set as the median value of the fuzzy intensity variation range, and is calculated using the following formula: ,in and These are the set minimum and maximum blur intensity thresholds, respectively; The noise intensity coefficient characterizes the severity of random fluctuations in turbulence. Gaussian white noise, which follows a standard normal distribution with a mean of 0 and a variance of 1, introduces randomness to simulate the unpredictability of atmospheric flow.

[0041] The fuzzy intensity sequence generated by the above formula It is applied in real-time to image degradation processing, ensuring that the blur level of each captured image frame is different and statistically correlated. (Attached) Figure 6Five representative images generated by this dynamic process are shown, corresponding from left to right to the 1st, 101st, 201st, 301st, and 401st samples in the iteration process. Observation reveals that the image sharpness of the extended target (UAV) fluctuates randomly over time. This non-stationary input signal effectively tests the feature extraction stability of the target detection model under conditions of drastic changes in signal-to-noise ratio, as well as the robustness of the correction algorithm under dynamic interference. This dynamic modeling based on the OU process, compared to simple random noise superposition, effectively simulates the dynamic characteristics of real atmospheric turbulence, thus demonstrating that this system is not only suitable for static laboratory environments but also possesses the technical potential to cope with complex outdoor weather conditions.

[0042] See attached document Figure 2 After image acquisition is completed, the controller executes step S2, which involves processing the distorted image using a deep learning neural network object detection model. Figure 2 This is a network structure diagram of the deep learning neural network object detection model (YOLOv11) in this embodiment of the invention; where input is the input terminal, used to input the collected light intensity distribution image; Backbone is the backbone network, used for feature extraction, extracting multi-scale feature maps; Neck is the neck component, used for feature fusion, aggregating deep semantics and shallow details; Head is the head component, used for predicting output, containing the detection head; Conv is a convolutional layer; C3k2, an improved cross-stage local network module, used to enhance feature extraction; SPPF is spatial pyramid pooling, used to fuse multi-scale features; C2PSA is a feature module with position-sensitive attention; Upsample: upsampling layer (used to enlarge the feature map size); Contact is used for channel fusion of feature maps; Detect is the detection head, outputting bounding boxes, presence probability, and class probability.

[0043] This invention utilizes deep learning technology to extract features from extended targets. In this embodiment, the YOLOv11 algorithm is used as an example of the deep learning neural network target detection model. However, the scope of this invention is not limited to the YOLO series models. Other mainstream target detection networks with similar feature extraction functions, such as SSD (SingleShot MultiBox Detector), Faster R-CNN, and RetinaNet, can also be applied to the control method of this invention after being trained on an extended target dataset. The architecture of YOLOv11 has been specifically optimized to meet the feature extraction requirements of extended targets. This model mainly consists of three core components: the backbone, the neck, and the head, which together complete the inference process from the original pixel input to the target confidence output.

[0044] The backbone network, acting as the model's front-end feature extractor, is responsible for receiving the raw RGB image data captured by the camera. This part employs a convolutional neural network (CNN) structure, transforming high-dimensional image pixels into abstract multi-scale feature maps through layer-by-layer convolution operations and downsampling. The backbone network can extract low-level features such as edges and textures, as well as high-level features such as shape and semantics from the image. Even when the input image suffers from severe atmospheric turbulence distortion, it can still effectively capture the key visual structural information of the extended target (drone).

[0045] The neck component, located between the backbone network and the head component, serves as an intermediate processing stage for feature fusion and enhancement. This component introduces mechanisms such as Feature Pyramid Network (FPN) or Path Aggregation Network (PAN), utilizing dedicated connection layers to aggregate feature maps of different scales output by the backbone network. By fusing deep semantic information with shallow detail information, the neck component enhances the model's ability to perceive targets of different sizes. Particularly for expanded targets that deform or become blurred in turbulent flow, it effectively preserves their key features and prevents the loss of information about small targets due to downsampling.

[0046] The head component, as the model's output, undertakes the final prediction task. Based on the refined feature map output by the neck component, this component performs target localization and classification. It contains multiple detector heads, each performing dense predictions on feature maps at different scales, generating a tensor output containing bounding box coordinates, target class probabilities, and target presence probabilities. The design of the head component directly determines the accuracy of the model's output confidence score and is a crucial foundation for subsequent performance metrics calculations.

[0047] To ensure the model can accurately identify specific extended targets (drones), this embodiment employs a targeted training strategy. The weight file loaded by the controller is trained on the publicly available Drone YOLO Detection Dataset, which contains approximately 4,000 drone images with different backgrounds and poses. In the training configuration, only a single category label, "quadcopter," is set to exclude interference from other irrelevant categories, allowing the model to focus on detecting a single target. Furthermore, transfer learning techniques are used, initializing network parameters using pre-trained weights YOLOv11n.pt. This approach leverages general feature representations learned from large-scale datasets, significantly accelerating convergence and improving the model's detection accuracy and generalization ability on specific tasks, enabling it to adapt to low-quality image inputs degraded by atmospheric turbulence. The specific implementation details of YOLOv11's convolutional layers and the backpropagation training algorithm are standard techniques in the field of computer vision and will not be elaborated upon here.

[0048] After the target detection model completes forward inference, the controller then executes step S3 to calculate the core performance metrics used to guide adaptive optics closed-loop correction. This invention innovatively proposes using the confidence score output by the deep learning model as the evaluation function, replacing the traditional Strel ratio (SR) or mean squared error (MSE).

[0049] The controller performs this calculation by running the core function `yolo_confidence(·)` in the Python code. This function receives the detection results of the current frame and extracts the detection bounding boxes with the highest confidence. If no target is detected in the current image, the function returns 0, corresponding to the extreme case where image distortion is so severe that target features are completely lost. When a target is detected, the system calculates the performance index according to the following formula. : ; In the above formula, This represents the probability of the target's existence, with a value ranging from 0 to 1. It reflects the model's confidence that the current detection box contains an object of interest. This represents the probability of the target category, with a value range of 0 to 1. It reflects the conditional probability that the object belongs to a preset category (in this embodiment, "quadcopter" drone) given that the object is confirmed to exist.

[0050] This composite indicator Not only are these detection parameters used in computer vision, but in this system they are also used as a basis for quantifying optical imaging quality. When atmospheric turbulence causes severe wavefront distortion, the image becomes blurred and the edges become diffuse, making it difficult for the YOLO model's feature extraction network to capture effective geometric features, thus hindering... and The numerical value decreases, ultimately resulting in a lower confidence level in the output. Conversely, as the corrector effectively compensates for wavefront distortion, image sharpness improves, feature details are restored, and the model's confidence in target recognition increases, leading to improved performance. It then rises.

[0051] By establishing this mapping relationship between image quality and recognition confidence, this invention solves the problem of the disconnect between traditional optical evaluation metrics and practical application requirements. Traditional metrics such as Strel ratio are mainly designed for point light sources, and in extended target scenes, they often exhibit the phenomenon that correction converges but target details remain blurry and difficult to identify. Confidence metrics, on the other hand... By using the recognition probability of machine vision tasks as the optimization target, a high degree of consistency between the adaptive optics correction direction and the back-end intelligent processing requirements is ensured.

[0052] This design enables the system to complete closed-loop optimization based solely on image content feedback without the need for a wavefront sensor, greatly simplifying the system's optical path structure.

[0053] See attached document Figure 1 After obtaining the target recognition confidence score, the controller executes step S4, which involves closed-loop control of the corrector based on the Stochastic Parallel Gradient Descent (SPGD) algorithm. This process constructs a direct feedback loop based on the image and confidence score, completely replacing the function of the wavefront sensor in traditional adaptive optics systems and solving the problem of difficult wavefront detection under low signal-to-noise ratio conditions for extended targets.

[0054] Specifically, the controller first bases its control voltage vector on... Generate a set of random voltage perturbation vectors that satisfy a Bernoulli distribution. The dimension of this perturbation vector is consistent with the number of drive units in the corrector, corresponding to the 61 control channels of the 61-unit deformable mirror in this embodiment. This is to estimate the performance indicators. The controller employs a two-sided disturbance detection strategy relative to the gradient direction of the control voltage.

[0055] First, the controller will apply a positive disturbance voltage. An application is made to the corrector, causing a small, random deformation of the deformable mirror surface. After the light beam is reflected by this deformed mirror, its phase changes, and the camera subsequently captures the corresponding positive perturbation image. The controller uses the YOLOv11 model to infer from this image and calculates the positive performance index. .

[0056] Subsequently, the controller will apply a negative disturbance voltage. An error is applied to the corrector, causing a slight, reverse deformation of the deformable mirror surface. The camera then re-exposures to capture the corresponding negative perturbation image, and the controller similarly calculates the negative performance index. .

[0057] During this process, the system actually acquires multiple frames of image data, including both positive and negative disturbance images. The controller then uses positive performance indicators... With negative performance indicators The difference This is used to approximate the gradient information of the performance function in the current high-dimensional voltage space. This mechanism of estimating the gradient by applying a small tentative perturbation and observing the system output response (confidence change) allows the system to determine the voltage adjustment direction required to optimize image quality without resolving the specific distribution of the light wave phase, thus realizing blind optimization control under wavefront-free conditions.

[0058] See attached document Figure 1After completing gradient estimation, the controller converts the semantic-level indicators output by the deep learning model into electrical signals that drive the physical optical devices using specific mathematical rules. This process achieves deep integration of the algorithm and the model, enabling the neural network's recognition capabilities to directly affect the physical modulation of the light wave phase. The specific voltage update process is executed according to the following formula: ; In the formula, Representing the The control voltage vector applied to each drive unit of the corrector during the next iteration. For the first The base voltage vector at the next iteration. The difference between the performance indices under the positive and negative perturbations calculated above represents the magnitude and direction of the gradient. The vector is the applied random voltage disturbance. In this embodiment, the elements of the vector follow a Bernoulli distribution and have values ​​of ±0.5V. The learning rate is a scalar gain coefficient that controls the voltage update step size.

[0059] Learning rate With disturbance amplitude The parameter settings are crucial to the system's convergence performance. If the parameters are set too small, the voltage correction in each iteration is negligible, requiring a large number of iterations to reach the optimal state, and the response speed cannot meet the real-time correction requirements. If the parameters are set too large, the voltage update amplitude is drastic, easily causing the performance indicators to oscillate significantly or even diverge near the extreme points, making stable convergence impossible. After experimental optimization, this embodiment sets the learning rate... Setting it to 0.5 will reduce the disturbance amplitude. The amplitude is set to 0.5. The controller calculates the updated voltage vector. The signal is then sent to a high-voltage amplifier driver, which applies the corresponding analog voltage to the actuator of the deformable mirror, thereby changing the mirror's surface shape. This parameter allows the system to recover from severe wavefront distortion to a clear imaging state within approximately 500 iterations. Through this iterative update mechanism, the controller continuously corrects the corrector's mirror surface shape, improving the system's performance indicators. The gradual increase until it reaches its maximum value and stabilizes indicates that the optical path phase distortion has been effectively compensated. The specific mathematical derivation of the SPGD algorithm is a standard technique in numerical optimization and will not be elaborated upon here.

[0060] See attached document Figure 4 and attached Figure 5 During the iterative update process of step S5, which is continuously executed by the controller, the system is subjected to static Gaussian random phase perturbation ( The correction effect under the ) environment was quantitatively and qualitatively verified.

[0061] In static scene testing, the target recognition confidence level in the initial stage (approximately the first 50 iterations) after controller initialization is achieved. The value is in the low region close to 0, corresponding to the corrector not yet effectively compensating for wavefront distortion, and the image acquired by the camera is still extremely blurry. As the SPGD algorithm continues to update the voltage vector according to Equation 2, the performance index curve shows a significant upward trend.

[0062] As attached Figure 5 As shown in the figure, the title "YOLO confidence vs SPGD iteration" indicates the relationship between YOLO confidence and the number of SPGD iterations; "YOLO max confidence" (Y-axis) represents the maximum confidence of YOLO (i.e., the performance metric); and "Iteration" (X-axis) represents the number of iterations.

[0063] When the iterations reach approximately 150, the confidence level is... A steep ascent indicates that the optimization algorithm has found the correct gradient descent direction, and the root mean square error of the wavefront phase has significantly decreased. When the number of iterations reaches approximately 350 to 500, the confidence curve gradually flattens and stabilizes in a high-value region above 0.8, indicating that the system has entered a convergent state. At this point, refer to the attached... Figure 4 The corrected extended target (drone) image exhibits clear geometric contours and rich surface texture details, compared to the attached... Figure 3 The blurred images shown present a stark contrast. This demonstrates that a confidence-based closed-loop control strategy can effectively guide the deformable mirror to reconstruct the phase of the light wave, achieving high-resolution imaging.

[0064] See attached document Figure 7 To verify the system's adaptability to time-varying atmospheric turbulence, the controller was tested in a dynamic aberration environment generated by the OU process (corresponding to the attached...). Figure 6 The closed-loop algorithm is run on the randomly degraded image sequence shown. In this scenario, the blurriness of the input image... Random fluctuations over time lead to changes in system performance metrics. It is no longer a smooth, monotonous upward curve, but rather shows an upward trend accompanied by local oscillations. Figure 7 In the diagram, YOLO conf (legend) represents the YOLO confidence level; YOLO maxconfidence (Y-axis) represents the maximum confidence level of YOLO; and Iteration (X-axis) represents the number of iterations.

[0065] observe Figure 7It can be seen that despite the continuous changes in external turbulence intensity, the controller can still maintain the mean of the confidence index at a high level through rapid voltage iteration updates. In the early stage of iteration, the system quickly responds to large aberration changes, raising the confidence level from a low point. After entering the stable tracking stage, although the confidence curve exhibits sawtooth fluctuations due to environmental disturbances, it remains above the effective recognition threshold without divergence or frame drops. This convergence characteristic proves that the system not only possesses static aberration correction capabilities but also dynamic tracking bandwidth, enabling it to maintain a dynamic balance of imaging quality through continuous gradient estimation and voltage correction under conditions where both the extended target and atmospheric turbulence parameters change simultaneously.

[0066] See attached document Figure 8 To further verify the superiority of the deep learning-based performance metrics proposed in this invention in practical applications, this embodiment introduces image grayscale variance, commonly used in traditional adaptive optics, as a performance metric for the comparison group under the same hardware optical path and atmospheric turbulence simulation conditions. The controller uses image variance and target recognition confidence as the optimization objective functions of the SPGD algorithm, respectively, and records and compares their convergence trajectories during the closed-loop correction process. Figure 8 In the figure, "Comparison of Variance (normalized) and YOLO Confidence" (title) compares the normalized image variance with the YOLO target recognition confidence; "Value (YOLO:0-1, Variance:normalized 0-1)" (Y-axis) represents the index value (where the YOLO confidence is the original value from 0 to 1, and the image variance is normalized and mapped to the 0-1 range); "Iteration" (X-axis) represents the number of iterations; "YOLO Confidence" (legend) represents the YOLO target recognition confidence; and "Variance" (legend) represents the image variance.

[0067] In the appendix Figure 8 In the graph, the horizontal axis represents the number of iterations, and the vertical axis represents the normalized performance index value. Observing the comparison curves, it can be seen that the convergence curve based on image variance (usually showing a relatively gentle upward trend) grows slowly in the early stages of iteration. This is because the variance index mainly measures the dispersion of pixel gray levels across the entire image. For the expanded target, it cannot distinguish between effective edge texture information and background noise or artifacts, resulting in an unclear gradient estimation direction in the multidimensional voltage space and low search efficiency. Furthermore, the variance curve exhibits significant high-frequency fluctuations during its ascent and remains highly volatile even after reaching a steady state, indicating that the index is sensitive to environmental noise and the system's stability is poor.

[0068] In contrast, the target recognition confidence level proposed in this invention... The convergence curve exhibits significant performance advantages. First, in terms of convergence speed, the confidence index shows a very steep upward slope in the first 200 iterations after startup, enabling it to quickly escape the fuzzy state and reach the neighborhood of the optimal solution with fewer iterations. This is due to the nonlinear extraction capability of the deep learning model for target features, making the performance index more sensitive to the improvement of wavefront aberrations, thus providing a steeper and more accurate gradient guide for the SPGD algorithm.

[0069] Secondly, regarding curve smoothness and stability, the confidence curve exhibits smoother convergence with minimal local oscillations. This is because the YOLOv11 model, when calculating confidence, naturally filters out high-frequency random noise in the image through the hierarchical structure of the convolutional neural network, responding only to pixel changes related to the UAV's structural features (such as rotor and fuselage shape). This strong correlation with target semantic features allows the control system to ignore interference from non-target areas and focus on optimizing the imaging quality of the target itself.

[0070] The final experimental results clearly demonstrate that, compared to traditional image variance metrics, the wavefront-detection-free adaptive optics control method, which uses target recognition confidence as a performance metric, not only improves image clarity but, more importantly, maximizes the probability of machine vision recognition. This metric design ensures that the corrected image can directly serve backend intelligent analysis tasks, solving the technical problem that traditional metrics may encounter in extended target correction where "the image becomes brighter or the contrast is improved, but the target is still difficult to identify." This strongly supports the advanced nature and practicality of this invention for extended target imaging correction in complex atmospheric environments.

Claims

1. A control method for a wavefront-free adaptive optics system, characterized in that, The optical system includes an extended target, a combined optical path beam-shrinking system, a corrector, a high-voltage amplifier driver, an imaging lens, a camera, and a controller; The control method includes the following steps: Step S1: The camera acquires an image of the light intensity distribution of the extended target after being beam-shrinked by the combined optical path beam-shrinking system and reflected by the corrector. The light intensity distribution image contains wavefront distortion information after transmission through atmospheric turbulence, and the light intensity distribution image is transmitted to the controller. Step S2: The controller inputs the light intensity distribution image into a preset deep learning neural network target detection model. The deep learning neural network target detection model performs feature extraction and localization on the extended target and outputs a detection result containing target detection boxes and probability information. Step S3: The controller extracts the detection box information with the highest confidence from the detection results, calculates the target recognition confidence based on the detection box information, and uses the target recognition confidence as a performance index of adaptive optics closed-loop control. Step S4: The controller applies a bilateral random perturbation voltage to the corrector based on the stochastic parallel gradient descent algorithm, and estimates the gradient direction of the performance index based on the change of the performance index corresponding to the image acquired by the camera under the perturbation. Step S5: The controller updates the control voltage vector of the corrector according to the estimated gradient direction, and drives the actuator of the corrector to change the mirror surface shape through the high voltage amplifier driver to compensate for the wavefront distortion until the performance index reaches the convergence state.

2. The control method for a wavefront-free adaptive optics system according to claim 1, characterized in that, In step S1, the static aberration environment corresponding to the wavefront distortion information is constructed by a Gaussian random phase perturbation model. The Gaussian random phase perturbation model is generated based on the Gaussian blur kernel function and uses the standard deviation parameter to quantify the blur degree of the image.

3. The control method for a wavefront-free adaptive optics system according to claim 1, characterized in that, In step S1, the dynamic aberration environment corresponding to the wavefront distortion information is constructed through the Ornstein-Uhlenbeck process, wherein the evolution of the dynamic fuzz intensity follows a stochastic differential equation, which is determined by the regression rate, the expected fuzz mean, and the noise intensity coefficient, in order to simulate the mean regression characteristics and stochastic fluctuation characteristics of the turbulence intensity.

4. The control method for a wavefront-free adaptive optics system according to claim 1, characterized in that, In step S2, the deep learning neural network target detection model includes a backbone network, a neck component, and a head component connected in sequence; The backbone network adopts a convolutional neural network structure to extract multi-scale feature maps from the light intensity distribution image; the neck component adopts a feature fusion network structure to aggregate the deep semantic information and shallow detail information output by the backbone network; the head component contains multiple detection heads to generate the detection result based on the aggregated feature map, and the detection result includes bounding box coordinates, target presence probability, and target category probability.

5. The control method for a wavefront-free adaptive optics system according to claim 4, characterized in that, In step S3, the value of the performance index is equal to the product of the target existence probability and the target category probability; wherein, the target existence probability is used to characterize the degree of confidence of the deep learning neural network target detection model in that the current detection box contains an object of interest, and the target category probability characterizes the conditional probability that the object belongs to a preset category given that the object exists.

6. The control method for a wavefront-free adaptive optics system according to claim 1, characterized in that, Step S4 specifically includes: Generate a set of random voltage perturbation vectors that satisfy a Bernoulli distribution, wherein the dimension of the random voltage perturbation vectors is the same as the number of driving units of the corrector; The base voltage vector at the current moment is added to the random voltage disturbance vector to obtain the positive disturbance voltage, which is then applied to the corrector to acquire a positive disturbance image and calculate the positive performance index. The negative perturbation voltage is obtained by subtracting the base voltage vector from the random voltage perturbation vector, and is applied to the corrector. A negative perturbation image is acquired and a negative performance index is calculated. The difference between the positive performance index and the negative performance index is calculated, and the difference is used to characterize the gradient information of the performance index in the current voltage space.

7. The control method for a wavefront-free adaptive optics system according to claim 6, characterized in that, In step S5, the update rule for the control voltage vector is as follows: The updated control voltage vector is obtained by adding the base voltage vector to the correction term; the correction term is determined by the product of the learning rate, the difference, and the random voltage perturbation vector, wherein the learning rate is used to control the step size of the voltage update.

8. The control method for a wavefront-free adaptive optics system according to claim 1, characterized in that, In step S3, if the deep learning neural network target detection model does not contain any target detection box in the detection result at the current time, the controller directly sets the target recognition confidence value to zero, which is used to characterize the state where the target features are completely lost due to severe image distortion.

9. The control method for a wavefront-free adaptive optics system according to claim 4, characterized in that, The deep learning neural network object detection model is trained using a transfer learning strategy. The weights are initialized using a pre-trained weight file, and the training dataset is a public dataset containing drone images. The class labels during training are limited to a single class.

10. The control method for a wavefront-free adaptive optics system according to claim 1, characterized in that, In step S5, the controller is connected to the high-voltage amplifier driver via a D / A conversion interface. The high-voltage amplifier driver is used to receive and amplify the analog voltage signal from the controller. The amplified voltage signal is then transmitted to each actuator electrode of the corrector to drive the mirror to deform and modulate the beam phase.

Images