An adaptive optical wavefront reconstruction method based on quantum-inspired feature screening

Abstract

The application discloses an adaptive optical wavefront reconstruction method based on quantum heuristic feature screening, and relates to the technical field of wavefront reconstruction. The method comprises the following steps: acquiring a light spot image sample and corresponding wavefront phase data, calculating Zernike coefficient labels to construct a wavefront reconstruction training data set; dividing the light spot image sample into multiple candidate image blocks, establishing a correlation evaluation model between the candidate image blocks and the Zernike coefficient labels and a redundancy evaluation model between the candidate image blocks, and constructing a quadratic unconstrained binary optimization model; obtaining a target image block set by using a quantum heuristic optimization algorithm, generating a space mask, and screening the light spot image; training a convolutional neural network by using the light spot image processed by the mask, and completing wavefront reconstruction by using the trained convolutional neural network.

Description

Technical Field

[0001] This invention relates to the field of wavefront reconstruction technology, and more specifically to an adaptive optical wavefront reconstruction method based on quantum-inspired feature screening. Background Technology

[0002] Adaptive optics systems typically utilize Shaker-Hartmann wavefront sensors to acquire wavefront distortion information and drive a wavefront corrector to compensate for optical distortion based on the wavefront reconstruction results. With the development of deep learning technology, methods that use convolutional neural networks to directly predict Zernike coefficients from spot images to complete wavefront reconstruction are gradually being applied in the field of adaptive optics. This can reduce the traditional centroid extraction and wavefront slope calculation processes and improve wavefront reconstruction efficiency.

[0003] However, existing wavefront reconstruction methods based on convolutional neural networks typically use complete spot images as network input. These complete spot images contain a large number of background regions and redundant regions that contribute little to wavefront reconstruction, increasing the computational cost of the convolutional neural network and reducing model inference efficiency. Furthermore, most existing feature selection methods rely on a single evaluation metric to select input regions, failing to comprehensively consider the correlation between image regions and wavefront parameters, information redundancy between image regions, and the constraint of the number of input regions. This results in the selected input regions failing to balance effective information preservation and computational complexity reduction. In addition, existing feature dimensionality reduction methods often lack a correspondence between features and their spatial locations in the original spot image, hindering the formation of a fixed input region selection strategy and failing to meet the application requirements of real-time wavefront reconstruction.

[0004] Therefore, there is a need to provide an adaptive optical wavefront reconstruction method that can achieve effective region selection of spot images, reduce redundant input while retaining the effective information required for wavefront reconstruction, and provide support for improving wavefront reconstruction efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive optics wavefront reconstruction method based on quantum-inspired feature selection. First, multiple candidate image blocks are obtained by dividing the spot image. A quadratic unconstrained binary optimization model is established by considering the correlation between candidate image blocks and Zernike coefficient labels, the feature redundancy among candidate image blocks, and the constraint of the number of target image blocks. A quantum-inspired optimization algorithm is used to determine the target image blocks that contribute significantly to wavefront reconstruction. A fixed spatial mask is then constructed and used to perform spatial region selection on the spot image to be reconstructed. This reduces redundant input regions while retaining the effective information required for wavefront reconstruction. The selected spot image is then input into a convolutional neural network for Zernike coefficient prediction, thus achieving wavefront reconstruction. By combining spatial region optimization with convolutional neural network prediction, the convolutional neural network can learn features from image regions highly correlated with wavefront reconstruction, providing more effective input data and reducing the participation of invalid regions in the calculation. This provides technical support for improving wavefront reconstruction efficiency and establishes a fixed input region selection strategy corresponding to the spatial location of the original spot image, facilitating its application in adaptive optics systems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides an adaptive optical wavefront reconstruction method based on quantum-inspired feature screening, the method comprising: Acquire light spot image samples and corresponding wavefront phase data collected by the Shaker-Hartmann wavefront sensor, and calculate the corresponding Zernike coefficient labels based on the wavefront phase data to construct a wavefront reconstruction training dataset. The light spot image is divided into multiple candidate image blocks, and the image features corresponding to each candidate image block are extracted; Based on the correlation between the image features of each candidate image patch and the Zernike coefficient labels, as well as the feature similarity among the candidate image patches, a quadratic unconstrained binary optimization model is established for selecting target image patches; the quadratic unconstrained binary optimization model is expressed as: ; Where, x i A binary variable r represents whether the i-th candidate image patch is selected. i c represents the correlation evaluation value between the i-th candidate image patch and the multi-order Zernike coefficient label. ij Let represent the redundancy evaluation value between the i-th candidate image block and the j-th candidate image block, K represent the preset number of target image blocks, and α, β and γ represent the weight coefficients of the correlation term, the redundancy penalty term and the quantity constraint term, respectively. A quantum-inspired optimization algorithm is used to solve a quadratic unconstrained binary optimization model to obtain a set of target image patches, and a spatial mask is constructed based on the set of target image patches. Spatial regions are filtered using a spatial mask to obtain a masked spot image; The masked spot image is input into a convolutional neural network for wavefront parameter prediction, and wavefront reconstruction is completed based on the predicted Zernike coefficients.

[0007] In some embodiments, a wavefront reconstruction training dataset is constructed, including: An atmospheric turbulence model is established based on preset atmospheric turbulence parameters; Wavefront phase screens are generated based on atmospheric turbulence models; The wavefront phase screen is fitted using Zernike polynomials to obtain the corresponding Zernike coefficient labels. The imaging process of the Shak-Hartmann wavefront sensor is simulated using a wavefront phase screen to obtain the corresponding light spot image. Preset noise is added to the spot image to form training samples corresponding to the Zernike coefficient labels.

[0008] In some embodiments, establishing a quadratic unconstrained binary optimization model includes: The light spot image is divided into blocks according to a preset size, and each image block is used as a candidate image block. Calculate the correlation evaluation value between each candidate image patch and the Zernike coefficient label; Calculate the redundancy evaluation value between any two candidate image patches; A quadratic unconstrained binary optimization objective function is established based on the correlation evaluation value, the redundancy evaluation value, and the preset number of target image blocks.

[0009] In some embodiments, solving a quadratic unconstrained binary optimization model using a quantum-inspired optimization algorithm includes: The quadratic unconstrained binary optimization model is solved iteratively using simulated annealing, quantum annealing, or quantum-inspired optimization algorithms. The selection status of each candidate image patch is determined based on the optimization results; A spatial mask corresponding to the spot image is created based on the selected state, wherein the region corresponding to the target image block is the preserved region, and the remaining regions are the suppressed regions.

[0010] In some embodiments, spatial region filtering is performed on the image of the light spot to be reconstructed using a spatial mask, including: Map the spatial mask onto the light spot image to be reconstructed; The image information within the area corresponding to the space mask is preserved; Suppress or nullify the image information within the corresponding suppression area of ​​the space mask; Output the image of the light spot after masking.

[0011] In some embodiments, inputting the masked spot image into a convolutional neural network for wavefront parameter prediction includes: Feature extraction of the masked spot image is performed using a convolutional neural network. Establish a mapping relationship between image features and Zernike coefficients based on the extracted image features; Output the Zernike coefficient prediction results; The wavefront phase is reconstructed based on the Zernike coefficient prediction results.

[0012] Secondly, the present invention also provides an adaptive optical wavefront reconstruction device based on quantum-inspired feature screening, the device comprising: The data construction module is used to acquire light spot image samples and corresponding wavefront phase data collected by the Shaker-Hartmann wavefront sensor, and calculate the corresponding Zernike coefficient labels based on the wavefront phase data to construct the wavefront reconstruction training dataset. The feature extraction module is used to divide the spot image into multiple candidate image blocks and extract the image features corresponding to each candidate image block; The model building module is used to build a quadratic unconstrained binary optimization model for screening target image patches based on the correlation between the image features of each candidate image patch and the Zernike coefficient label, as well as the feature similarity between each candidate image patch. The model solving module is used to solve the quadratic unconstrained binary optimization model using a quantum-inspired optimization algorithm to obtain the target image patch set, and to construct a spatial mask based on the target image patch set; The region filtering module is used to perform spatial region filtering on the light spot image to be reconstructed using a spatial mask, so as to obtain the light spot image after mask processing. The wavefront reconstruction module is used to input the masked spot image into a convolutional neural network to predict wavefront parameters and complete wavefront reconstruction based on the predicted Zernike coefficients.

[0013] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the adaptive optical wavefront reconstruction method based on quantum-inspired feature screening provided in the first aspect.

[0014] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the adaptive optical wavefront reconstruction method based on quantum-inspired feature screening provided in the first aspect.

[0015] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the adaptive optical wavefront reconstruction method based on quantum-inspired feature screening provided in the first aspect.

[0016] The beneficial effects of this invention are as follows: This invention provides an adaptive optics wavefront reconstruction method based on quantum-inspired feature selection. First, multiple candidate image blocks are obtained by dividing the spot image. A quadratic unconstrained binary optimization model is established by considering the correlation between candidate image blocks and Zernike coefficient labels, the feature redundancy among candidate image blocks, and the constraint of the number of target image blocks. A quantum-inspired optimization algorithm is used to determine the target image blocks that contribute significantly to wavefront reconstruction. A fixed spatial mask is then constructed, and the spatial mask is used to perform spatial region selection on the spot image to be reconstructed. This reduces redundant input regions while retaining the effective information required for wavefront reconstruction. The selected spot image is then input into a convolutional neural network for Zernike coefficient prediction, thus achieving wavefront reconstruction. By combining spatial region optimization with convolutional neural network prediction, the convolutional neural network can learn features from image regions highly correlated with wavefront reconstruction, providing more effective input data for wavefront reconstruction and reducing the participation of invalid regions in the calculation. This provides technical support for improving wavefront reconstruction efficiency. Simultaneously, a fixed input region selection strategy corresponding to the spatial location of the original spot image is formed, facilitating its application in adaptive optics systems.

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of an adaptive optical wavefront reconstruction method based on quantum-inspired feature screening, as shown in an embodiment of the present invention. Figure 2 This is an example image of a Shaker-Hartmann noisy spot, as shown in an embodiment of the present invention. Figure 3 This is a schematic diagram of spot image segmentation and candidate image block division according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the spatial mask generation and mask processing effect according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a convolutional neural network wavefront reconstruction model structure according to an embodiment of the present invention; Figure 6 This is a schematic diagram of wavefront reconstruction results based on predicted Zernike coefficients according to an embodiment of the present invention; Figure 7This is a schematic diagram of an adaptive optical wavefront reconstruction device based on quantum-inspired feature screening, according to an embodiment of the present invention. Figure 8 This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0019] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0020] It should be noted that references to "an embodiment," "embodiment," "example embodiment," etc., in this specification refer to the described embodiment including specific features, structures, or characteristics; however, not every embodiment must include these specific features, structures, or characteristics. Furthermore, such expressions do not refer to the same embodiment. Moreover, when describing specific features, structures, or characteristics in conjunction with embodiments, whether or not explicitly described, it is indicated that incorporating such features, structures, or characteristics into other embodiments is within the knowledge of those skilled in the art.

[0021] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0022] Example 1 like Figure 1 As shown, this embodiment provides an adaptive optics wavefront reconstruction method based on quantum-inspired feature selection and convolutional neural networks. This method is applied to adaptive optics systems to rapidly reconstruct the corresponding wavefront phase from a light spot image acquired by a Shaker-Hartmann wavefront sensor, providing wavefront information for subsequent wavefront correction. The method mainly includes the construction of a wavefront reconstruction training dataset, candidate image patch selection, construction of a quadratic unconstrained binary optimization model, quantum-inspired optimization solution, spatial mask generation, and convolutional neural network training and wavefront reconstruction. Each step is executed sequentially according to the data flow, ultimately establishing a mapping relationship between the light spot image and the wavefront phase, achieving rapid wavefront reconstruction.

[0023] S101: Obtain light spot image samples and corresponding wavefront phase data collected by the Shaker-Hartmann wavefront sensor. Calculate the corresponding Zernike coefficient labels based on the wavefront phase data to construct a wavefront reconstruction training dataset.

[0024] The Shack-Hartmann Wavefront Sensor (SHWFS) is a wavefront detection device used to detect wavefront distortion information of light waves. It primarily uses a microlens array to divide the incident light beam into multiple sub-apertures, each forming a corresponding light spot. When the light beam is affected by factors such as atmospheric turbulence, resulting in wavefront distortion, the center positions of the light spots formed by different sub-apertures will shift to varying degrees. By analyzing the changes in the positions of each light spot, the local slope changes of the wavefront can be reflected. Therefore, the Shack-Hartmann Wavefront Sensor can acquire light spot images reflecting the wavefront distortion characteristics, such as... Figure 2 As shown, Figure 2 Example image of a Shaker-Hartmann image with noisy spot.

[0025] In this embodiment, the light spot image samples refer to the Shak-Hartmann light spot images used to train the convolutional neural network. Each light spot image sample corresponds to a unique set of wavefront phase data. Wavefront phase data refers to the phase distribution of light waves on the pupil plane after propagation. It can truly reflect the atmospheric turbulence disturbances experienced by the current light beam and is also an important basis for subsequent calculation of Zernike coefficient labels and evaluation of wavefront reconstruction accuracy.

[0026] It should be noted that "acquisition" in this embodiment can refer to either data collected by the actual adaptive optics system or simulation generation based on the system model. First, an atmospheric turbulence model is established to simulate the beam propagation process. Corresponding wavefront phase data is generated through the atmospheric turbulence model. Then, the imaging process of the Shaker-Hartmann wavefront sensor is simulated based on the generated wavefront phase data to obtain corresponding spot image samples, thereby establishing a one-to-one correspondence between the spot image samples and the wavefront phase data.

[0027] The atmospheric turbulence model is used to simulate wavefront distortion caused by random changes in refractive index during beam propagation in the atmosphere, enabling the generated data to cover different levels of atmospheric turbulence conditions and improving the representativeness of the training data. After obtaining the wavefront phase data, it is input into the Shak-Hartmann wavefront sensor model to simulate the sampling process of each microlens on the sub-aperture wavefront, generating corresponding Shak-Hartmann spot images. Simultaneously, to more closely approximate the imaging environment in actual adaptive optics systems, noise is further added when generating the spot images to obtain noisy spot images. Using noisy spot images as network training data enables the trained convolutional neural network to have better environmental adaptability, improving its wavefront reconstruction performance in practical applications.

[0028] After obtaining the light spot image sample and the corresponding wavefront phase data, the corresponding Zernike coefficient label is calculated based on the wavefront phase data.

[0029] The Zernike coefficients are a set of expanded coefficients obtained by expanding the wavefront phase using Zernike orthogonal polynomials. Different orders of Zernike coefficients correspond to different types of wavefront aberrations, and all Zernike coefficients collectively characterize the overall wavefront morphology. Therefore, in this embodiment, the wavefront phase data is not directly used as the supervision label for the convolutional neural network. Instead, the corresponding multi-order Zernike coefficients are first calculated based on the wavefront phase data, and the calculated multi-order Zernike coefficients are used as the network training labels, enabling the convolutional neural network to learn the mapping relationship between spot image samples and Zernike coefficients.

[0030] Each spot image sample corresponds to a set of multi-order Zernike coefficient labels. Therefore, after calculating the Zernike coefficients, a one-to-one correspondence is established between each spot image sample and its corresponding Zernike coefficient label, ultimately forming the wavefront reconstruction training dataset. The wavefront reconstruction training dataset includes multiple spot image samples and their corresponding Zernike coefficient labels, providing a unified data foundation for subsequent candidate image patch segmentation, quantum-inspired feature selection, convolutional neural network training, and wavefront reconstruction.

[0031] It should be noted that the wavefront reconstruction training dataset in this embodiment can be constructed offline in one go, or it can be expanded and updated based on newly added training samples. When new spot image samples and corresponding wavefront phase data are acquired, the corresponding Zernike coefficient labels can be calculated using the same method as described above, and added to the existing wavefront reconstruction training dataset, thereby further enriching the training samples and improving the learning ability of the convolutional neural network for different wavefront distortion conditions.

[0032] In this embodiment, following the dataset generation process, spot image samples and corresponding wavefront phase data collected by a Shaker-Hartmann wavefront sensor are obtained. Multi-order Zernike coefficient labels are then calculated based on the wavefront phase data to construct a wavefront reconstruction training dataset. On one hand, this ensures that each spot image sample has a corresponding supervisory label, providing standard training samples for subsequent convolutional neural network training. On the other hand, by using noisy spot images to construct the training dataset, the trained wavefront reconstruction model can adapt to more complex real-world imaging environments, thereby improving the accuracy and stability of subsequent wavefront reconstruction.

[0033] S102, the light spot image is divided into multiple candidate image blocks, and the image features corresponding to each candidate image block are extracted.

[0034] In this context, candidate image blocks refer to multiple local image regions segmented from a single spot image sample according to a preset division method. Each candidate image block contains a portion of the spot information from the original spot image and serves as the basic processing unit for subsequent spatial region optimization. Since the spot image acquired by the Shaker-Hartmann wavefront sensor contains spots formed by multiple sub-apertures, and the spot regions corresponding to different spatial locations are not entirely of equal importance for wavefront reconstruction, directly inputting the entire spot image into a convolutional neural network for training would not only increase the network's computational load but also allow a large number of low-contribution background or low-information regions to participate in feature learning, affecting the network's training efficiency. Therefore, this embodiment first divides the complete spot image into multiple candidate image blocks, and then selects regions that contribute significantly to wavefront reconstruction from these candidate image blocks, thereby reducing the amount of data that the subsequent network needs to process.

[0035] It should be noted that the image segmentation in this embodiment is performed while maintaining the spatial relationship of the original spot image. That is, image segmentation only changes the data organization and does not change the wavefront information contained in the original image. Each candidate image block corresponds to a unique spatial location in the original spot image. Therefore, after the target image block is subsequently selected, its positional relationship in the original spot image can still be accurately recovered, providing a positional basis for the subsequent generation of the spatial mask.

[0036] like Figure 3 As shown, this embodiment uses a sliding window method to divide the spot image samples. A sliding window refers to setting an image window to a preset size and moving it sequentially along the horizontal and vertical axes of the spot image. Each movement extracts a corresponding region as a candidate image block, until the entire image has been traversed. Using a sliding window method ensures that candidate image blocks cover the entire spot image, allowing different spatial regions to participate in subsequent feature selection. It also maintains spatial continuity between candidate image blocks, which is beneficial for subsequent analysis of the contribution of different spatial regions to wavefront reconstruction.

[0037] Preferably, the window size and movement step size of the sliding window can be set according to the resolution of the spot image. When the window size is small, a larger number of candidate image patches can be obtained, thus allowing for a more detailed analysis of the importance of different local regions; when the window size is appropriately increased, the number of candidate image patches can be reduced, improving the efficiency of subsequent optimization solutions. In this embodiment, window partitioning parameters are preferably used to generate candidate image patches to ensure that the subsequent spatial mask can accurately cover the regions that contribute significantly to wavefront reconstruction.

[0038] After dividing the candidate image blocks, each candidate image block inherits the Zernike coefficient label corresponding to the original spot image. That is, all candidate image blocks from the same spot image sample correspond to the same set of Zernike coefficient labels. By establishing the correspondence between candidate image blocks and Zernike coefficient labels, it is possible to further evaluate the importance of different candidate image blocks for Zernike coefficient prediction, providing a supervised basis for quantum-inspired feature selection.

[0039] After obtaining all candidate image blocks, the corresponding image features are extracted for each candidate image block.

[0040] Image features refer to the data representation that describes the content of candidate image patches. They reflect the wavefront information contained in the current candidate image patch and provide a basis for evaluating the importance of the candidate image patch. Since different candidate image patches correspond to different spatial locations, their light spot morphology, grayscale distribution, and spatial structure are all different. Therefore, by extracting image features, the original image data can be converted into a data format that is easy to analyze later.

[0041] This embodiment obtains image features corresponding to each candidate image patch through the feature extraction process of the front end of the convolutional neural network. That is, it uses convolution operations to extract local features from the candidate image patches, so that the candidate image patches gradually form feature representations that can characterize local wavefront information, rather than manually designing feature parameters. Since convolution operations can automatically learn the image features contained in different spatial regions, they can better adapt to the changing patterns of spot images under different wavefront distortion conditions, thus improving the reliability of subsequent feature selection.

[0042] Furthermore, the image features corresponding to each candidate image patch not only retain the local spatial information in the original image, but also maintain the positional correspondence between the candidate image patches. Therefore, when establishing a secondary unconstrained binary optimization model, the image features and the spatial positions of the candidate image patches can be combined to uniformly evaluate each candidate image patch, and finally select the set of target image patches that can represent the main wavefront information.

[0043] In this embodiment, the spot image samples in the wavefront reconstruction training dataset are first divided into multiple candidate image blocks, and the image features corresponding to each candidate image block are extracted. This transforms the entire spot image into multiple independently analyzable local regions, providing a unified data foundation for subsequent establishment of a quadratic unconstrained binary optimization model and the use of quantum-inspired optimization algorithms to select target image blocks. Simultaneously, by maintaining the spatial correspondence between the candidate image blocks and the original spot image, accurate positional mapping is provided for subsequent spatial mask generation and convolutional neural network wavefront reconstruction, improving data consistency and processing efficiency throughout the entire wavefront reconstruction process.

[0044] After dividing each candidate image block, the image features corresponding to each candidate image block are extracted.

[0045] Image features refer to data representations that characterize the content of candidate image blocks. They reflect the spot information contained in the candidate image blocks and provide a data foundation for evaluating the importance of candidate image blocks to wavefront reconstruction. In this embodiment, the image features are derived from the image data contained in the candidate image blocks themselves, rather than constructing a new image description model. Therefore, each candidate image block can form corresponding image features after being divided, while maintaining the same spatial positional relationship as the original spot image.

[0046] In the specific implementation process, for each candidate image block, the image data within the corresponding region is read and saved as the image features of the corresponding candidate image block, so that each candidate image block forms corresponding image feature data. At the same time, the spatial location information corresponding to the candidate image block and the Zernike coefficient label corresponding to the light spot image sample are linked together to form complete candidate image block information, providing input data for subsequent quantum-inspired feature screening.

[0047] Since all candidate image patches originate from the same wavefront reconstruction training dataset, the image features corresponding to the candidate image patches not only retain the image information of the local spot region, but also inherit the supervision labels of the corresponding training samples. This enables subsequent analysis of the contribution of different candidate image patches to the wavefront reconstruction task based on the correspondence between image features and Zernike coefficient labels.

[0048] It should be noted that in this embodiment, image features mainly serve as information carriers for candidate image blocks. The focus is on completely preserving the image information of the corresponding regions of the candidate image blocks, rather than completing deep feature learning at this stage. Therefore, this embodiment does not require building a convolutional neural network model during the candidate image block segmentation stage, nor does it require completing network training in advance. Instead, when building a secondary unconstrained binary optimization model later, the image features corresponding to each candidate image block are used to further analyze the correlation between the candidate image block and the Zernike coefficient label, as well as the degree of information redundancy among the candidate image blocks, thereby achieving the selection of target image blocks.

[0049] Through the above method, this embodiment completes the candidate image block division and corresponding image feature extraction for each spot image sample in the wavefront reconstruction training dataset. This transforms the original spot image into multiple independently analyzable local regions while maintaining the spatial correspondence between each candidate image block and the original spot image. This provides a unified data foundation for subsequent establishment of a quadratic unconstrained binary optimization model and for spatial region optimization using quantum-inspired optimization algorithms. Furthermore, since the image features are directly derived from the candidate image blocks themselves, the wavefront information contained in the local spot regions can be fully preserved, avoiding the introduction of additional data processing before candidate image block selection. This simplifies the entire wavefront reconstruction process and maintains consistency with subsequent spatial mask generation and convolutional neural network training processes.

[0050] S103. Based on the correlation between the image features of each candidate image patch and the Zernike coefficient label, as well as the feature similarity between each candidate image patch, a quadratic unconstrained binary optimization model is established for screening target image patches.

[0051] The quadratic unconstrained binary optimization model is expressed as: ; Where, x i A binary variable r represents whether the i-th candidate image patch is selected. i c represents the correlation evaluation value between the i-th candidate image patch and the multi-order Zernike coefficient label. ij Let represent the redundancy evaluation value between the i-th candidate image block and the j-th candidate image block, K represent the preset number of target image blocks, and α, β and γ represent the weight coefficients of the correlation term, the redundancy penalty term and the quantity constraint term, respectively.

[0052] The correlation evaluation model is used to assess the degree of association between the image information contained in candidate image patches and the Zernike coefficient labels, that is, to evaluate the importance of the current candidate image patch for the wavefront reconstruction task. Since the wavefront information contained in the spot regions corresponding to different spatial locations varies, the contribution of different candidate image patches to Zernike coefficient prediction also differs. The purpose of establishing the correlation evaluation model is to identify candidate image patches that can better reflect the wavefront distortion characteristics from all candidate image patches, providing an evaluation basis for subsequent spatial region optimization.

[0053] In the specific implementation process, a correspondence is established between the image features corresponding to each candidate image patch and the Zernike coefficient labels corresponding to their respective spot image samples. Based on the correlation between image features and Zernike coefficient labels, a correlation evaluation is performed on each candidate image patch to obtain the correlation evaluation result for each candidate image patch. Each candidate image patch corresponds to a correlation evaluation value, which characterizes the degree to which the candidate image patch contributes to Zernike coefficient prediction. A larger correlation evaluation value indicates that the information contained in the candidate image patch is more important for wavefront reconstruction; a smaller correlation evaluation value indicates that the candidate image patch contributes relatively less to wavefront reconstruction.

[0054] Relying solely on relevance evaluation results to select candidate image patches can easily lead to situations where multiple selected candidate image patches contain similar image information. That is, while multiple candidate image patches may have high relevance, there is significant information overlap between them, resulting in redundancy in the data input to the convolutional neural network. Therefore, this embodiment further establishes a redundancy evaluation model.

[0055] The redundancy evaluation model is used to assess the degree of information redundancy between different candidate image patches, that is, to evaluate whether two candidate image patches contain similar image information. When two candidate image patches have a high degree of information redundancy, even if they participate in subsequent wavefront reconstruction simultaneously, it is difficult to provide more new and effective information, and it will only increase the data processing load of the convolutional neural network. Therefore, while ensuring the accuracy of wavefront reconstruction, we should try to minimize the number of candidate image patches with high redundancy information that enter the final selection result at the same time.

[0056] In the specific implementation process, based on the image features corresponding to each candidate image patch, the information overlap between any two candidate image patches is evaluated to obtain the redundancy evaluation result between the candidate image patches. For any set of candidate image patches, a corresponding redundancy evaluation value can be obtained. When the redundancy evaluation value is small, it indicates that the two candidate image patches can provide more complementary information; when the redundancy evaluation value is large, it indicates that the two candidate image patches contain more identical or similar information, and do not need to be retained simultaneously in the subsequent spatial region optimization process.

[0057] After obtaining the correlation evaluation results for each candidate image patch and the redundancy evaluation results between candidate image patches, a quadratic unconstrained binary optimization model is constructed based on the correlation evaluation model and the redundancy evaluation model.

[0058] The Quadratic Unconstrained Binary Optimization (QUBO) model is a mathematical model that uses binary variables to describe combinatorial optimization problems. It can unify multiple optimization objectives into a single objective function and use optimization algorithms to search for the optimal solution to that function. In this embodiment, each candidate image patch corresponds to a binary decision variable. When the corresponding binary decision variable takes the first state, it indicates that the corresponding candidate image patch is selected; when the corresponding binary decision variable takes the second state, it indicates that the corresponding candidate image patch is discarded. Therefore, all binary decision variables jointly determine which candidate image patches are ultimately retained.

[0059] The quadratic unconstrained binary optimization model constructed in this embodiment comprehensively considers the importance of candidate image patches and the degree of information redundancy among them. Specifically, the correlation evaluation model corresponds to the correlation optimization term in the quadratic unconstrained binary optimization model, used to increase the overall information content of the final selected target image patches; the redundancy evaluation model corresponds to the redundancy constraint term in the quadratic unconstrained binary optimization model, used to reduce the degree of information redundancy among target image patches, enabling the final selected target image patches to cover more different types of wavefront information.

[0060] Furthermore, to ensure that the number of target image patches obtained after final selection meets the input requirements of the subsequent convolutional neural network, this embodiment also establishes a corresponding quantity constraint based on the preset number of target image patches. This ensures that the quadratic unconstrained binary optimization model not only considers correlation and redundancy during the optimization process but also guarantees that the final set of target image patches meets the preset quantity requirement. In this way, it can avoid the increase in computational complexity of the convolutional neural network due to an excessive number of target image patches, and also avoid the loss of effective wavefront information due to an insufficient number of target image patches.

[0061] like Figure 4 As shown, after the quadratic unconstrained binary optimization model is established, the candidate image patch selection problem is uniformly transformed into a combinatorial optimization problem. That is, by optimizing the combination of binary decision variables, the optimal spatial region combination that simultaneously satisfies the constraints of high correlation, low redundancy, and the number of target image patches is satisfied is found, providing a unified mathematical model for subsequent quantum-inspired optimization algorithm solutions.

[0062] In this embodiment, the first step is to establish a correlation evaluation model based on the image features corresponding to the candidate image blocks and the Zernike coefficient labels. Then, a redundancy evaluation model is established based on the image features between the candidate image blocks. Furthermore, a quadratic unconstrained binary optimization model is constructed, which transforms the candidate image block selection from the traditional method of selecting spatial regions based on experience to an automatic selection method based on mathematical optimization models. This provides a unified optimization objective for the subsequent quantum-inspired optimization algorithm to automatically search the target image block set, and also provides a theoretical basis for the generation of fixed spatial masks.

[0063] S104. A quantum-inspired optimization algorithm is used to solve the quadratic unconstrained binary optimization model to obtain a set of target image patches, and a spatial mask is constructed based on the set of target image patches.

[0064] After completing the construction of the quadratic unconstrained binary optimization model, in order to determine the target spatial region that will ultimately participate in wavefront reconstruction from all candidate image patches, this embodiment uses a quantum-inspired optimization algorithm to solve the quadratic unconstrained binary optimization model and obtain the target image patch set.

[0065] The quadratic unconstrained binary optimization model only establishes the optimization objective corresponding to the candidate image patch selection and the constraint relationship between each candidate image patch. Which candidate image patches should be retained still needs to be determined through optimization. Therefore, this embodiment further employs a quantum-inspired optimization algorithm to solve the quadratic unconstrained binary optimization model.

[0066] Quantum-inspired optimization algorithms are a class of combinatorial optimization algorithms designed with the optimization ideas of quantum computing. They can use the concept of quantum evolution to search for complex combinatorial optimization problems and quickly obtain optimization results within a large solution space. In this embodiment, the quantum-inspired optimization algorithm is used as a solution tool for a quadratic unconstrained binary optimization model. Its main function is to search for combinations of target image patches that satisfy the optimization conditions among all candidate image patches based on the established optimization objective, rather than directly processing the spot image.

[0067] In the specific implementation process, the established quadratic unconstrained binary optimization model is input into the quantum-inspired optimization algorithm, and the binary decision variables corresponding to the candidate image patches are used as the optimization objects. After the optimization begins, the quantum-inspired optimization algorithm continuously searches and updates each combination of binary decision variables based on the correlation evaluation results, redundancy evaluation results, and the constraint of the number of target image patches in the quadratic unconstrained binary optimization model, so that the optimization objective gradually converges.

[0068] As the optimization process continues, the binary decision variables corresponding to each candidate image patch gradually stabilize. When the quantum-inspired optimization algorithm reaches the preset termination condition, it outputs the final optimization result. The selection of each candidate image patch is determined based on the final values ​​of the binary decision variables, thus obtaining the target image patch set.

[0069] The target image blocks in the target image block set are not selected independently, but are obtained by overall optimization based on comprehensive consideration of the correlation and redundancy among all candidate image blocks. Therefore, the target image blocks can maintain a high degree of information complementarity and can better retain the effective information related to wavefront reconstruction in the spot image.

[0070] Since the target image patch set is derived from the global optimization results of the entire wavefront reconstruction training dataset, this embodiment can automatically obtain a spatial region that is more suitable for the wavefront reconstruction task compared to the method of determining the region of interest based on human experience, thus providing target region information for subsequent spatial mask generation.

[0071] In this embodiment, the quantum-inspired optimization algorithm is used to solve the quadratic unconstrained binary optimization model and obtain the target image block set. This transforms the candidate image block selection process from manual experience selection to automatic optimization, improving the rationality of spatial region selection and providing a unified target region for subsequent spatial mask generation.

[0072] After obtaining the target image patch set, in order to apply the optimization results obtained by the quantum-inspired optimization algorithm to the subsequent convolutional neural network training and wavefront reconstruction process, this embodiment constructs a spatial mask based on the target image patch set, and uses the spatial mask to perform spatial region filtering on the spot image samples to obtain the masked spot image samples.

[0073] The spatial mask is a spatial mapping matrix used to identify the location of the target region in the spot image. It has the same spatial size as the original spot image. The spatial mask can clearly distinguish which regions participate in the subsequent wavefront reconstruction and which regions do not participate in the subsequent wavefront reconstruction.

[0074] like Figure 4 As shown, in this embodiment, based on the spatial position of the target image patch set in the original spot image, each target image patch is remapped to its corresponding position in the original spot image, and a spatial mask is generated accordingly. The position corresponding to the target image patch is marked as the reserved region, while the remaining regions are marked as non-reserved regions, thus forming a complete spatial mask.

[0075] Since the spatial mask originates from the set of target image patches obtained by optimizing the entire training dataset using a quantum-inspired optimization algorithm, this spatial mask is a fixed spatial mask. That is, the same spatial mask is used in both the training stage of the convolutional neural network and the subsequent wavefront reconstruction stage, without having to recalculate the target region for each spot image.

[0076] S105, Spatial region filtering is performed on the light spot image to be reconstructed using a spatial mask to obtain the light spot image after mask processing.

[0077] After constructing the spatial mask, it is applied to the spot image samples in the wavefront reconstruction training dataset. Spatial regions are then filtered based on the positional relationships corresponding to the spatial mask. Specifically, image data within the target region corresponding to the spatial mask is retained, while non-target regions are masked. This ensures that the final masked spot image samples primarily retain information relevant to wavefront reconstruction.

[0078] After spatial masking, the target region in the spot image retains the original image information, while a large number of background regions and spatial regions that contribute less no longer participate in the subsequent training of the convolutional neural network, thereby reducing the amount of data that the convolutional neural network needs to process.

[0079] Since the spatial mask remains consistent between the training and inference phases, the spatial features learned during convolutional neural network training can be directly applied to the actual wavefront reconstruction process, improving the consistency between model training and practical application.

[0080] In this embodiment, a fixed spatial mask is generated based on the target image block set, and the spatial mask is used to complete the spatial region selection of the light spot image. This allows the convolutional neural network to learn only the regions that contribute significantly to wavefront reconstruction, thereby reducing redundant information in the network input and providing more effective input data for subsequent convolutional neural network training.

[0081] S106, input the masked spot image into the convolutional neural network to predict wavefront parameters, and complete wavefront reconstruction based on the predicted Zernike coefficients.

[0082] After obtaining the masked spot image samples, the masked spot image samples are input into the convolutional neural network for training. The trained convolutional neural network then outputs the corresponding Zernike coefficients based on the spot image to be reconstructed, thereby completing the wavefront reconstruction.

[0083] Convolutional neural networks are used to establish a mapping relationship between masked spot image samples and Zernike coefficient labels, enabling the network to predict the corresponding Zernike coefficients based on the input spot image and achieve wavefront reconstruction.

[0084] like Figure 5 As shown, the masked spot image samples are used as input to a convolutional neural network (CNN), and the corresponding Zernike coefficient labels from the wavefront reconstruction training dataset are used as supervision labels to train the CNN, enabling it to continuously learn the correspondence between the spot images and the Zernike coefficients. Once the network training reaches the preset requirements, a wavefront reconstruction model is obtained.

[0085] After completing network training, when wavefront reconstruction is required, the Shaker-Hartmann spot image corresponding to the beam to be reconstructed is first obtained. Then, the spatial region of the spot image to be reconstructed is filtered using the aforementioned fixed spatial mask to obtain the masked spot image.

[0086] Subsequently, the masked image of the light spot to be reconstructed is input into the trained convolutional neural network, which then outputs the corresponding Zernike coefficient prediction results.

[0087] Since the Zernike coefficients can characterize wavefront aberrations, the corresponding wavefront phase can be reconstructed based on the Zernike coefficients output by the convolutional neural network, thus completing wavefront reconstruction.

[0088] In practical adaptive optics systems, the reconstructed wavefront information can be further used in wavefront correction devices. Based on the reconstructed wavefront distortion information, corresponding control quantities are generated to drive wavefront correction devices such as deformable mirrors, thereby achieving real-time compensation for wavefront distortion and improving the imaging quality of the adaptive optics system.

[0089] Through the above steps, this embodiment completes the entire process from constructing the wavefront reconstruction training dataset, dividing candidate image blocks, quantum-inspired feature selection, spatial mask generation, convolutional neural network training to wavefront reconstruction. It realizes the automatic selection of spatial regions that contribute significantly to wavefront reconstruction using quantum-inspired optimization algorithms, and the rapid completion of wavefront reconstruction by combining convolutional neural networks. This provides an adaptive optics system with a wavefront reconstruction method that balances wavefront reconstruction accuracy and efficiency.

[0090] Example 2 To verify the effectiveness of the adaptive optical wavefront reconstruction method based on quantum-inspired feature screening proposed in this application, the overall experimental process of this embodiment mainly includes steps such as training data construction, spatial region optimization, convolutional neural network training, and wavefront reconstruction performance evaluation.

[0091] This embodiment uses the training dataset constructed in Embodiment 1 as the source of experimental data. First, atmospheric turbulence phase screens with different degrees of wavefront distortion are generated based on preset atmospheric turbulence parameters. Then, corresponding light spot images are generated using a Shaker-Hartmann wavefront sensor simulation model. Simultaneously, Zernike polynomials are used to fit each wavefront phase screen, obtaining the corresponding multi-order Zernike coefficients as supervision labels. To make the training data more closely resemble the actual engineering application environment, detector noise and environmental noise are further added during the generation of light spot images, enabling the network training process to cover wavefront reconstruction tasks under different noise conditions.

[0092] Subsequently, spatial region optimization processing was performed on all training samples using the method of Example 1.

[0093] To verify the impact of spatial region optimization strategies on wavefront reconstruction performance, this embodiment sets up two sets of comparative experiments. The control group uses the traditional convolutional neural network (CNN) wavefront reconstruction method, which directly inputs the complete Shaker-Hartmann spot image into the CNN for training and prediction. The experimental group uses the method described in this application, which first uses a quantum-inspired optimization algorithm to generate a fixed spatial mask, performs spatial region filtering on the spot image, and then inputs the masked spot image into the CNN for training and prediction. Both sets of experiments use the same dataset, the same CNN structure, the same training parameters, and the same optimizer; only the input data processing method differs to ensure the comparability of the experimental results.

[0094] After network training, the trained model is used to predict the Zernike coefficients for each spot image in the test set, and the wavefront phase is reconstructed based on the predicted Zernike coefficients. Simultaneously, the true wavefront is reconstructed based on the true Zernike coefficients corresponding to the test samples. The predicted wavefront is then compared with the true wavefront, and the root mean square error (RMSE) is used as the main evaluation metric to evaluate the wavefront reconstruction performance of the two methods. A smaller RMSE indicates a smaller deviation between the predicted and true wavefronts, and higher wavefront reconstruction accuracy.

[0095] Furthermore, to evaluate the improvement effect of the proposed method on network inference efficiency, this embodiment further statistically analyzes the average inference time required for the two methods to complete single-frame wavefront reconstruction and calculates the number of effective pixels in the network input participating in the convolution calculation. Since this application first utilizes a spatial mask to remove a large number of background regions and image regions with low contribution, the actual effective input that the convolutional neural network needs to process is significantly reduced, thereby further reducing the network inference time.

[0096] Experimental results are as follows Figure 6As shown in the figure, the experimental results demonstrate that, using the method described in this application, the convolutional neural network primarily extracts features from the spatial regions closely related to wavefront reconstruction. The predicted Zernike coefficients more accurately reflect the actual wavefront distortion, and the reconstructed predicted wavefront exhibits high consistency with the true wavefront. Furthermore, compared to methods that directly use complete spot images as network input, this method effectively reduces the data scale involved in the convolutional neural network's computation while maintaining high wavefront reconstruction accuracy, thus lowering network inference time and improving wavefront reconstruction efficiency. This better meets the application requirements of closed-loop adaptive optics systems for real-time wavefront reconstruction.

[0097] Further analysis of the experimental results reveals that the main reason for the aforementioned effectiveness of the proposed method lies in the fact that it does not employ traditional random cropping or empirical region selection methods. Instead, it comprehensively considers the correlation between each candidate image patch and the Zernike coefficient, the degree of information redundancy among candidate image patches, and the constraint of the number of input regions. By establishing a quadratic unconstrained binary optimization model, the optimal set of target image patches is determined. This ensures that the retained spatial regions contain more information related to wavefront reconstruction while avoiding multiple adjacent regions repeatedly expressing the same wavefront features. Furthermore, after generating a fixed spatial mask based on the target image patches, all subsequent images of light spots to be reconstructed can be preprocessed using the same spatial region selection strategy, eliminating the need to recalculate the target region for each image. This not only reduces the computational overhead in the online inference stage but also ensures the consistency of network input and improves the stability of the entire wavefront reconstruction process.

[0098] In summary, this embodiment verifies that the adaptive optical wavefront reconstruction method based on quantum-inspired feature screening proposed in this application can effectively optimize the selection of the effective spatial region of the Shaker-Hartmann spot image. It reduces the input redundancy of the convolutional neural network while preserving the effective information of wavefront reconstruction as much as possible, improves the wavefront reconstruction efficiency, and maintains good wavefront reconstruction accuracy, providing experimental verification for the effectiveness of the technical solution of this application.

[0099] Based on the same inventive concept, this application also provides an adaptive optics wavefront reconstruction device based on quantum-inspired feature selection for implementing the adaptive optics wavefront reconstruction method based on quantum-inspired feature selection described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the adaptive optics wavefront reconstruction device based on quantum-inspired feature selection provided below can be found in the limitations of the adaptive optics wavefront reconstruction method based on quantum-inspired feature selection described above, and will not be repeated here.

[0100] In one embodiment, such as Figure 7As shown, an adaptive optical wavefront reconstruction device based on quantum-inspired feature screening is provided. The device includes: The data construction module 30 is used to acquire light spot image samples and corresponding wavefront phase data collected by the Shaker-Hartmann wavefront sensor, and calculate the corresponding Zernike coefficient labels based on the wavefront phase data to construct a wavefront reconstruction training dataset. The feature extraction module 31 is used to divide the light spot image into multiple candidate image blocks and extract the image features corresponding to each candidate image block; The model building module 32 is used to build a quadratic unconstrained binary optimization model for screening target image blocks based on the correlation between the image features of each candidate image block and the Zernike coefficient label, as well as the feature similarity between each candidate image block. The model solving module 33 is used to solve the quadratic unconstrained binary optimization model using a quantum-inspired optimization algorithm to obtain a set of target image patches and to construct a spatial mask based on the set of target image patches. The region filtering module 34 is used to perform spatial region filtering on the light spot image to be reconstructed using a spatial mask, so as to obtain the light spot image after mask processing. The wavefront reconstruction module 35 is used to input the masked spot image into the convolutional neural network to predict the wavefront parameters and complete the wavefront reconstruction based on the predicted Zernike coefficients.

[0101] This application also provides an electronic device, in some embodiments, referring to... Figure 8 As shown, the electronic device 700 includes an input unit 710, a memory 720, a processor 730, and an output unit 740. The memory 720 stores program instructions that can be executed on the processor 730. The processor 730 can execute the adaptive optical wavefront reconstruction method and / or technical solution based on quantum-inspired feature screening as described in the foregoing embodiments by calling the program instructions. The electronic device 700 can be a mobile terminal device such as a mobile phone or a computer.

[0102] Furthermore, embodiments of this application also provide a computer-readable storage medium for storing a computer program that executes an adaptive optical wavefront reconstruction method based on quantum-inspired feature screening. For example, computer program instructions, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. The program instructions that invoke the methods of this application may be stored in a fixed or removable storage medium, and / or transmitted via data streams in broadcast or other signal carrying media, and / or stored in a storage medium that operates according to the program instructions.

[0103] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0104] The technical features of the above embodiments can be arbitrarily integrated. For the sake of brevity, not all possible integrations of the technical features in the above embodiments are described. However, as long as the integration of these technical features does not contradict each other, they should be considered to be within the scope of this specification.

[0105] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. An adaptive optical wavefront reconstruction method based on quantum-inspired feature screening, characterized in that, The method includes: Acquire light spot image samples and corresponding wavefront phase data collected by the Shaker-Hartmann wavefront sensor, and calculate the corresponding Zernike coefficient labels based on the wavefront phase data to construct a wavefront reconstruction training dataset. The light spot image is divided into multiple candidate image blocks, and the image features corresponding to each candidate image block are extracted; Based on the correlation between the image features of each candidate image patch and the Zernike coefficient label, as well as the feature similarity between the candidate image patches, a quadratic unconstrained binary optimization model is established for filtering target image patches; the quadratic unconstrained binary optimization model is expressed as: ； Where, x i A binary variable r represents whether the i-th candidate image patch is selected. i c represents the correlation evaluation value between the i-th candidate image patch and the multi-order Zernike coefficient label. ij Let represent the redundancy evaluation value between the i-th candidate image block and the j-th candidate image block, K represent the preset number of target image blocks, and α, β and γ represent the weight coefficients of the correlation term, the redundancy penalty term and the quantity constraint term, respectively. The quadratic unconstrained binary optimization model is solved using a quantum-inspired optimization algorithm to obtain a set of target image patches, and a spatial mask is constructed based on the set of target image patches. The spatial mask is used to perform spatial region filtering on the light spot image to be reconstructed, resulting in a light spot image after mask processing. The masked spot image is input into a convolutional neural network for wavefront parameter prediction, and wavefront reconstruction is completed based on the predicted Zernike coefficients.

2. The adaptive optical wavefront reconstruction method based on quantum-inspired feature screening as described in claim 1, characterized in that, The construction of the wavefront reconstruction training dataset includes: An atmospheric turbulence model is established based on preset atmospheric turbulence parameters; A wavefront phase screen is generated based on the atmospheric turbulence model. The wavefront phase screen is fitted using Zernike polynomials to obtain the corresponding Zernike coefficient labels; The imaging process of the Shaker-Hartmann wavefront sensor is simulated by the wavefront phase screen to obtain the corresponding spot image. Preset noise is added to the light spot image to form training samples corresponding to the Zernike coefficient labels.

3. The adaptive optical wavefront reconstruction method based on quantum-inspired feature screening as described in claim 1, characterized in that, The establishment of the quadratic unconstrained binary optimization model includes: The light spot image is divided into blocks according to a preset size, and each image block is used as a candidate image block. Calculate the correlation evaluation value between each candidate image patch and the Zernike coefficient label; Calculate the redundancy evaluation value between any two candidate image patches; A quadratic unconstrained binary optimization objective function is established based on the correlation evaluation value, the redundancy evaluation value, and the preset number of target image blocks.

4. The adaptive optical wavefront reconstruction method based on quantum-inspired feature screening as described in claim 3, characterized in that, The method of solving the quadratic unconstrained binary optimization model using a quantum-inspired optimization algorithm includes: The quadratic unconstrained binary optimization model is solved iteratively using simulated annealing, quantum annealing, or quantum-inspired optimization algorithms. The selection status of each candidate image patch is determined based on the optimization results; A spatial mask corresponding to the spot image is created based on the selected state, wherein the region corresponding to the target image block is the preserved region, and the remaining regions are the suppressed regions.

5. The adaptive optical wavefront reconstruction method based on quantum-inspired feature screening as described in claim 4, characterized in that, The spatial region selection of the light spot image to be reconstructed using a spatial mask includes: Map the spatial mask onto the light spot image to be reconstructed; The image information within the area corresponding to the space mask is preserved; Suppress or nullify the image information within the corresponding suppression area of ​​the space mask; Output the image of the light spot after masking.

6. The adaptive optical wavefront reconstruction method based on quantum-inspired feature screening as described in claim 5, characterized in that, The step of inputting the masked spot image into a convolutional neural network for wavefront parameter prediction includes: Feature extraction of the masked spot image is performed using a convolutional neural network. Establish a mapping relationship between image features and Zernike coefficients based on the extracted image features; Output the Zernike coefficient prediction results; The wavefront phase is reconstructed based on the Zernike coefficient prediction results.

7. An adaptive optical wavefront reconstruction device based on quantum-inspired feature screening, characterized in that, The device includes: The data construction module is used to acquire light spot image samples and corresponding wavefront phase data collected by the Shaker-Hartmann wavefront sensor, and calculate the corresponding Zernike coefficient labels based on the wavefront phase data to construct a wavefront reconstruction training dataset. The feature extraction module is used to divide the light spot image into multiple candidate image blocks and extract the image features corresponding to each candidate image block; The model building module is used to build a quadratic unconstrained binary optimization model for screening target image blocks based on the correlation between the image features of each candidate image block and the Zernike coefficient label, as well as the feature similarity between each candidate image block. The model solving module is used to solve the quadratic unconstrained binary optimization model using a quantum-inspired optimization algorithm to obtain a set of target image patches, and to construct a spatial mask based on the set of target image patches. The region filtering module is used to perform spatial region filtering on the light spot image to be reconstructed using the spatial mask, so as to obtain the light spot image after mask processing. The wavefront reconstruction module is used to input the masked spot image into a convolutional neural network to predict wavefront parameters and complete wavefront reconstruction based on the predicted Zernike coefficients.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the adaptive optical wavefront reconstruction method based on quantum-inspired feature screening as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the adaptive optical wavefront reconstruction method based on quantum-inspired feature screening as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the adaptive optical wavefront reconstruction method based on quantum-inspired feature screening as described in any one of claims 1 to 6.

Images