Atmospheric turbulence wavefront reconstruction method based on self-attention convolutional neural network
By combining a self-attention convolutional neural network with a Transformer structure, the high cost and complexity of wavefront sensor systems in atmospheric turbulence wavefront detection are solved, achieving high-precision wavefront reconstruction suitable for various detection scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHWESTERN POLYTECHNICAL UNIV
- Filing Date
- 2026-05-14
- Publication Date
- 2026-07-31
AI Technical Summary
In atmospheric turbulence wavefront detection, existing technologies suffer from high cost and complexity of wavefront sensor systems, and deep learning-based methods are insufficient in terms of accuracy and data set requirements, especially when dealing with extended targets.
A self-attention convolutional neural network-based approach was adopted. By constructing an optical path for atmospheric turbulence experiments to collect data and combining it with a deep neural network with a Transformer structure, wavefront reconstruction was achieved using a single intensity map. This approach avoids dependence on wavefront sensors and improves reconstruction accuracy.
It achieves high-precision wavefront reconstruction on point targets and extended targets, reduces hardware costs and system complexity, improves the generalization and reconstruction accuracy of the dataset, and reduces the dataset requirements.
Smart Images
Figure CN122492880A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of atmospheric turbulence wavefront detection technology, and specifically to a method for reconstructing atmospheric turbulence wavefronts based on self-attention convolutional neural networks. Background Technology
[0002] In the field of atmospheric turbulence wavefront sensing, detecting the distorted wavefront generated by turbulence has always been an important direction, as it determines the accuracy of subsequent wavefront compensation. Currently, the commonly used wavefront detection method is to use wavefront sensors. This method can achieve accurate wavefront reconstruction, but wavefront sensors mainly deal with point light source targets and have some limitations for extended targets. In addition, the system is relatively complex and costly.
[0003] Deep learning methods can be used to achieve rapid reconstruction of atmospheric turbulence wavefronts. Currently, deep learning-based wavefront detection methods are mainly divided into two types: wavefront detection based on wavefront sensor dot maps and wavefront-free detection based directly on intensity images. The former has higher accuracy, but it does not completely eliminate the limitations of wavefront sensors, while the latter still needs further optimization in terms of accuracy and dataset requirements.
[0004] To address the current limitations in accuracy and other issues associated with deep learning, several methods exist for improvement. For example, inputting a pair of in-focus and out-of-focus intensity maps captured simultaneously into the network can achieve more accurate feature extraction. However, this inadvertently increases the complexity of the dataset, inevitably introducing errors in actual captured data. Another approach is to combine neural networks with physical models. While this can increase the physical interpretability of the results to some extent, establishing an accurate physical model to combine with deep learning is very difficult when dealing with real turbulence.
[0005] To overcome the aforementioned limitations, this invention proposes a deep learning method combining a self-attention mechanism with a Transformer structure. This method achieves accurate atmospheric turbulence wavefront reconstruction using only a single intensity map, and its effectiveness has been validated on both point and extended target data. Furthermore, this invention does not require a wavefront sensor, has low hardware requirements, and is suitable for most atmospheric turbulence wavefront detection scenarios. Summary of the Invention
[0006] This invention proposes a method for reconstructing wavefronts of atmospheric turbulence based on a self-attention convolutional neural network (PAN). By incorporating a Tranformer structure, the accuracy of wavefront reconstruction is significantly improved. The method uses intensity images acquired through an atmospheric turbulence experiment and their corresponding phase screen images to form a dataset. A pre-built PAN is then trained on this dataset. Subsequent input to the trained network only requires a single intensity image to quickly output the corresponding wavefront phase. Compared to existing methods, this invention has lower requirements for the dataset, requiring only intensity images and corresponding phase images. Furthermore, the atmospheric turbulence experiment optical path structure is simple yet functional, capable of acquiring both point target and extended target data.
[0007] The specific technical solution of this invention is as follows: A method for reconstructing atmospheric turbulence wavefronts based on self-attention convolutional neural networks, characterized by the following steps: S1. Construct an experimental optical path for atmospheric turbulence, collect point target data and extended target data respectively, add phase distortion using a spatial light modulator, and capture intensity images using a CCD to obtain turbulence disturbance images under two real conditions; S2. Establish a neural network model. This invention uses a deep neural network that combines a self-attention convolution module, which includes two parts: convolutional upsampling and convolutional downsampling. Each sampling layer is composed of multiple stacked Transformer Blocks, which realizes complex turbulence feature extraction and directly realizes the mapping at the wavefront image level. S3. Perform preprocessing of the dataset, divide the point target data and extended target data, crop the image into a suitable shape, and perform normalization to facilitate network training; S4. Train the network using the point target dataset and the extended target dataset separately, setting the learning rate to 1e-4, adjusting the number of Transformer Blocks per layer to 8, and setting the loss function to RMS Loss. Subsequently, test the network using both point target and extended target data, calculating the reconstructed turbulence wavefront against the wavefront loaded by the spatial light modulator to quantitatively determine the network's reconstruction capability. The atmospheric turbulence experimental optical path in step S1 is used for dataset acquisition, but it is not limited to this; simulation or other experimental optical paths can also be used to obtain the corresponding dataset.
[0008] The number of Transformer Blocks in each sampling layer of the network in step S2 can be changed according to different requirements for network speed and accuracy.
[0009] The image preprocessing in step S3 can be cropped into different sizes according to the distribution of turbulence and the needs of different datasets, which facilitates network training.
[0010] The parameters such as the loss function and learning rate in step S4 can be changed according to different datasets and are not limited to RMS Loss and 1e-4.
[0011] The network reconstruction accuracy judgment in step S4 can be extended to the calculation of losses such as SSIM and PSNR, depending on the requirements for data structure accuracy and other aspects.
[0012] The beneficial effects of this invention lie in addressing the limitations of traditional wavefront detection systems in terms of cost and system complexity, as well as the accuracy deficiencies of wavefront-free systems based on traditional CNN networks. It proposes a self-attention-based convolutional neural network for wavefront reconstruction of atmospheric turbulence. This invention can directly output the wavefront at the image level through network training, avoiding errors in reconstruction coefficient fitting. It significantly improves wavefront reconstruction accuracy while reducing dataset requirements, demonstrating superior performance on both point target data and extended target data, and exhibiting high generalization ability. Attached Figure Description
[0013] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 Optical path diagram for atmospheric turbulence data acquisition experiment provided in this embodiment of the invention; Figure 2 A structural diagram of a self-attention-based convolutional neural network and TransformerBlock provided in an embodiment of the present invention; Figure 3 The wavefront reconstruction results and error map of point target data provided in the embodiments of the present invention; Figure 4 The extended target data wavefront reconstruction results and error map provided in the embodiments of the present invention; Figure 1 In the experiment, the optical path for atmospheric turbulence data acquisition can only acquire point target data and extended target data separately to avoid data accuracy errors caused by simultaneous acquisition. Figure 2 The diagram shows the main network structure and the structure of the Transformer Block, where each EncoderBlock, Latent Block, and Decoder Block is composed of multiple stacked Transformer Blocks. Figure 3 , 4The image shows the reconstruction results and errors of the point target and the extended target, respectively. The units for both the wavefront image and the residual image are rad. The error is calculated as RMS loss. The wavelength used in the optical path for acquiring the corresponding data is 532nm. Detailed Implementation
[0014] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.
[0015] A method for reconstructing atmospheric turbulence wavefronts based on self-attention convolutional neural networks, characterized by the following steps: S1. Construct an experimental optical path for atmospheric turbulence, collect point target data and extended target data respectively, add phase distortion using a spatial light modulator, and capture intensity images using a CCD to obtain turbulence disturbance images under two real conditions; S2. Establish a neural network model. This invention uses a deep neural network that combines a self-attention convolution module, which includes two parts: convolutional upsampling and convolutional downsampling. Each sampling layer is composed of multiple stacked Transformer Blocks, which realizes complex turbulence feature extraction and directly realizes the mapping at the wavefront image level. S3. Perform preprocessing of the dataset, divide the point target data and extended target data, crop the image into a suitable shape, and perform normalization to facilitate network training; S4. Train the network using the point target dataset and the extended target dataset separately, setting the learning rate to 1e-4, adjusting the number of Transformer Blocks per layer to 8, and setting the loss function to RMS Loss. Subsequently, test the network using both point target and extended target data, calculating the reconstructed turbulence wavefront against the wavefront loaded by the spatial light modulator to quantitatively determine the network's reconstruction capability. The atmospheric turbulence experimental optical path in step S1 is used for dataset acquisition, but it is not limited to this; simulation or other experimental optical paths can also be used to obtain the corresponding dataset.
[0016] The number of Transformer Blocks in each sampling layer of the network in step S2 can be changed according to different requirements for network speed and accuracy.
[0017] The image preprocessing in step S3 can be cropped into different sizes according to the distribution of turbulence and the needs of different datasets, which facilitates network training.
[0018] The parameters such as the loss function and learning rate in step S4 can be changed according to different datasets and are not limited to RMS Loss and 1e-4.
[0019] The network reconstruction accuracy judgment in step S4 can be extended to the calculation of losses such as SSIM and PSNR, depending on the requirements for data structure accuracy and other aspects.
[0020] Example 1: The workflow of an atmospheric turbulence wavefront reconstruction method based on a self-attention convolutional neural network is as follows: As described in step S1, an atmospheric turbulence experimental optical path is first constructed for data acquisition. This optical path can acquire two types of data: point target and extended target. The point target image is acquired by CCD1 after wavefront distortion is added by a spatial light modulator, and the extended target image is acquired by CCD2 after wavefront distortion is added by a spatial light modulator. The intensity map and the wavefront image are compared to form a dataset.
[0021] As described in step S2, a neural network model is constructed. The main body of the network is a deep convolutional neural network, which includes convolutional downsampling and convolutional upsampling. Each sampling layer consists of multiple Transformer Blocks, as shown in the specific structure below. Figure 1 As shown, each sampling layer here consists of 8 Transformer Blocks, and the Latent Block also consists of 8 Transformer Blocks. The network uses multiple Transformer Blocks stacked together to achieve complex feature extraction.
[0022] As described in step S3, we preprocess the dataset of 3000 pairs of data into training, validation and test sets in a ratio of 8:1:1. Since the image proportion of the point target spot is small, we crop the spot to a size of 128×128 and expand the target to a size of 256×256 for subsequent network processing.
[0023] As described in step S4, the network is trained using the point target data and the extended target data respectively. The intensity image is used as the input, the wavefront phase image as the label, the optimizer is set to Adam, the loss function is set to RMS Loss, and the learning rate is set to 1e-4. Training is then performed, followed by testing. We randomly selected three sets of data from each of the point target and extended target test sets, and the corresponding Input, Output, GT, and Error are shown below. Figure 3 , 4 As shown, the error calculation here uses the RMS loss of the residual map to judge the reconstruction accuracy from the standard of numerical error, and quantifies the wavefront reconstruction capability of the network.
[0024] The hardware and software equipment used in the method of this invention are: Ubuntu 22.04.5 operating system, 8 NVIDIA GeForce RTX 3090 graphics cards, Python 3.10 programming language, PyTorch 2.4.0 deep learning framework, Visual Studio Code compilation environment, and MATLAB 2024a simulation platform.
[0025] In summary, this invention relates to an atmospheric turbulence wavefront reconstruction method based on a self-attention convolutional neural network, aiming to address the problems of high cost and complex system structure in current turbulence wavefront reconstruction using wavefront detection systems, and insufficient accuracy and dataset limitations encountered in atmospheric turbulence wavefront reconstruction without wavefront detection systems based on traditional CNNs. This invention utilizes a self-attention-based convolutional neural network, achieving accurate feature extraction through the unique self-attention mechanism of the Transformer Block. Accurate wavefront reconstruction can be achieved with only a single intensity map. Furthermore, this method demonstrates superior performance in wavefront reconstruction for both point targets and extended targets, maintaining a low RMS loss in the reconstruction results.
[0026] In practical applications, the method involved in this invention has high wavefront reconstruction accuracy. The reconstruction result is trained and tested using data collected by the optical path of the atmospheric turbulence experiment. The RMS loss between the reconstruction result and the true value remains at a small value, making it suitable for most wavefront-less detection systems.
Claims
1. A method for atmospheric turbulence wavefront reconstruction based on self-attention convolutional neural network, characterized in that The method includes the following steps: S1. Construct an experimental optical path for atmospheric turbulence, collect point target data and extended target data respectively, add phase distortion using a spatial light modulator, and capture intensity images using a CCD to obtain turbulence disturbance images under two real conditions; S2. Establish a neural network model. This invention uses a deep neural network that combines a self-attention convolution module, which includes two parts: convolutional upsampling and convolutional downsampling. Each sampling layer is composed of multiple stacked Transformer Blocks, which realizes complex turbulence feature extraction and directly realizes the mapping at the wavefront image level. S3. Perform preprocessing of the dataset, divide the point target data and extended target data, crop the image into a suitable shape, and perform normalization to facilitate network training; S4. Train the network by training the point target dataset and the extended target dataset separately. Set the learning rate to 1e-4, adjust the number of Transformer Blocks in each layer to 8, and set the loss function to RMS Loss. Then, test the network using the point target and extended target datasets respectively. Calculate the reconstructed turbulent wavefront with the wavefront loaded by the spatial light modulator to quantitatively determine the reconstruction capability of the network.
2. The method for reconstructing atmospheric turbulence wavefronts based on a self-attention convolutional neural network according to claim 1, characterized in that, In step S1, the atmospheric turbulence experimental optical path needs to collect point target data and extended target data separately to avoid errors caused by simultaneous collection.
3. The method for reconstructing atmospheric turbulence wavefronts based on a self-attention convolutional neural network according to claim 1, characterized in that, The number of Transformer Blocks in each sampling layer of the network model in step S2 can be changed according to different requirements of data accuracy and processing speed, so as to achieve feature extraction with different levels of accuracy.
4. A method for reconstructing atmospheric turbulence wavefronts based on a self-attention convolutional neural network according to claim 1, characterized in that, The core of the network model in step S2 is a deep convolutional network, which can be further extended to other network structures such as fully connected networks while retaining the TransformerBlock.
5. The method for reconstructing atmospheric turbulence wavefronts based on a self-attention convolutional neural network according to claim 1, characterized in that, The image cropping shape in step S3 can be changed according to different datasets, which facilitates subsequent network training and improves reconstruction accuracy.
6. The method for reconstructing atmospheric turbulence wavefronts based on a self-attention convolutional neural network according to claim 1, characterized in that, The determination of network reconstruction accuracy in step S4 can be extended to the calculation of losses such as SSIM and PSNR, depending on the requirements for data structure accuracy and other aspects.