CN-122023160-A - Unsupervised reconstruction method for self-adaptive optical image recovery

Abstract

The invention discloses an unsupervised reconstruction method for self-adaptive optical image restoration, which is applied to the technical field of astronomical observation and aims at the problem of imaging resolution reduction caused by atmospheric turbulence in foundation solar telescope observation in the prior art; the invention innovatively integrates multi-frame blind deconvolution into an unsupervised deep learning framework, constructs an encoder-decoder network to generate a potential clear image, constructs a depth priori condition network optimization fuzzy core, and realizes high-efficiency recovery by solving network parameters and multi-frame blind deconvolution problems through alternate iteration. The GCM model is introduced into the corrector to ensure stable convergence, and a BM3D algorithm is adopted in the denoising process. Compared with the prior art, the method has the characteristics of no need of a large amount of training data, strong generalization capability, high recovery quality, capability of processing non-equivalent motion sickness and multiband images, realization of complementary advantages of the traditional algorithm and deep learning, high calculation efficiency, strong practicability and the like, and provides a more advanced and reliable image processing technology for astronomical observation and other fields.

Inventors

• WANG SHUAI
• LUO JIRUN
• HAO XIAOYANG
• LEI JINPENG

Assignees

• 电子科技大学长三角研究院(衢州)

Dates

Publication Date

20260512

Application Date

20260202

Claims (5)

1. 1. An unsupervised reconstruction method for adaptive optical image restoration, comprising: S1, acquiring continuous multi-frame short exposure sun images, and carrying out fuzzy kernel initialization processing; S2, taking the continuous multi-frame short-exposure solar image and the initialized fuzzy core obtained in the step S1 as input of a multi-frame check blind deconvolution algorithm, taking the input image into account of an image priori to obtain an intermediate estimated image, obtaining an estimated image by the intermediate estimated image through a check process, and obtaining an estimated fuzzy core by calculating the input fuzzy core; S3, constructing a jump connection network and a fuzzy core network based on the encoder-decoder architecture, wherein an estimated image obtained after the iterative computation of the multi-frame check blind deconvolution algorithm in the step S2 is used as the input of the jump connection network based on the encoder-decoder architecture, and a plurality of estimated fuzzy cores obtained after the iterative computation of the multi-frame check blind deconvolution algorithm in the step S2 are used as the input of the fuzzy core network; Taking the output of the jump connection network based on the encoder-decoder architecture and the output of the fuzzy core network as the input of the multi-frame check blind deconvolution algorithm in the current iteration of the step, taking the output of the multi-frame check blind deconvolution algorithm in the current iteration of the step as the input of the jump connection network based on the encoder-decoder architecture and the fuzzy core network in the next iteration of the step, thereby completing the updating of network parameters, returning to the step S2 if the set iteration times are reached and the objective function is not converged, and inputting the output of the multi-frame check blind deconvolution algorithm in the last iteration of the step S2; S4, recovering a clear image by adopting a non-blind deconvolution algorithm according to the estimated fuzzy kernel output by the multi-frame verification blind deconvolution algorithm after the convergence of the objective function.
2. 2. An unsupervised reconstruction method for adaptive optical image restoration according to claim 1, wherein the fuzzy core network comprises a plurality of single core networks, each core network being a two-layer fully connected network defined by fcn () function, the number of single core networks in the fuzzy core network being the same as the number of frames of the consecutive multiframes in step S1.
3. 3. An unsupervised reconstruction method for adaptive optical image restoration according to claim 2, wherein the objective function expression is: ; Wherein G x is a hopped network based on an encoder-decoder architecture, G ki is an i-th core network, For the convolution symbol, y i is the input i frame short exposure sun image, f is the clear image to be estimated, z x and z ki are intermediate iteration variables, For the purpose of the L2 regularization, For a low-quality a priori regularization, Are real numbers greater than 0.
4. 4. The unsupervised reconstruction method for adaptive optical image restoration according to claim 3, wherein the implementation process of the step S1 is that a circular pupil matrix is generated through parameterized design, spectrum centering is achieved through double ifftshift, two-dimensional Fourier transform is calculated efficiently by combining with an ifft2 algorithm, and finally normalized fuzzy kernels are obtained.
5. 5. An unsupervised reconstruction method for adaptive optics image restoration according to claim 4, wherein in the iterative process of step S2, the intermediate estimated image is calculated by BM3D algorithm.

Description

Unsupervised reconstruction method for self-adaptive optical image recovery Technical Field The invention belongs to the technical field of astronomical observation, and particularly relates to a self-adaptive optical image restoration technology for solar observation. Background Astronomical observation plays a key guiding role in the fields of production activities, aerospace industry, climate monitoring work, cosmic cognition exploration and the like. In the field of sun physics research, real-time monitoring of solar activity expansion by means of a telescope with a large view field and high resolution is one of important research means. The foundation telescope is extremely easy to be interfered by stray light and atmospheric turbulence when observing the sun, so that serious wave front distortion is generated on solar rays, and the imaging resolution is greatly reduced. To acquire high spatial resolution images, ground-based solar telescopes typically employ Adaptive Optics (AO) and image post-hoc reconstruction techniques. However, in the process of actually developing high-resolution solar astronomical observation, in view of the limitations of the correction order of the AO system, the detection precision of the wavefront sensor, the bandwidth of the control loop and the hardware performance of the deformable mirror, even the AO system with the most excellent performance can achieve a resolution of only 20% of the diffraction limit of the foundation telescope in terms of theoretical level. Furthermore, AO correction tends to be local only. Therefore, in order to enable the AO corrected image to be reconstructed further until the diffraction limit is reached, so as to obtain a high resolution imaging containing more useful information, the application of AO image post-processing techniques is particularly necessary. Currently, AO image post-processing methods can be roughly divided into two main categories, traditional algorithms and deep learning methods. The traditional algorithm mainly comprises a blind deconvolution method, a phase difference method and a speckle reconstruction method. The deep learning method can be mainly divided into two types, i.e. supervised type and unsupervised type. Blind deconvolution (Blind Deconvolution, BD) can reconstruct an image and estimate the blur kernel by means of a small number of frames in the case of unknown blur kernel, and has flexibility, effectiveness and wider application range because no reference object and degradation function are needed. However, since it belongs to the pathological inverse solution, artifacts are extremely easily introduced. The phase difference method (PHASE DIVERSITY, PD) needs to acquire a plurality of images by means of a beam splitter so as to solve wave fronts and clear images, and has the advantages of overcoming aberration and enhancing high-frequency information. The speckle reconstruction method (Spackle) is mainly applied to post-processing of large-caliber foundation telescope images, but has obvious defects of huge calculated amount and large data demand, and when short exposure frames are acquired, a target is required to be in a stable state and is required to accord with atmospheric turbulence statistical information, and when the weather is dim, effective analysis is difficult due to insufficient brightness. With the development of algorithm theory and the continuous improvement of hardware computing speed and performance, the deep learning-based method is in the brand-new angle in recent years and is rapidly developed. Multiple teams such as photoelectric stations in the Chinese academy of sciences have great achievements in AO image processing research, including application of multiple deep learning methods such as full convolutional networks, EDDNN, recurrent-DNN, CSSTN, cycleGAN architectures and the like. The current unsupervised deep learning method is less studied. The solar image is recovered by CycleGAN model in 2019 of Tai primitive university Gu Peng and the like, and belongs to a generalized unsupervised recovery algorithm. The physical loss function was constructed in 2021 by asensio Ramos et al, and the wavefront was estimated and the image was restored unsupervised using a convolutional network and a GRU unit. 2023. Self2Self unsupervised network denoising is constructed by U-Net at Nanjing university of the year. Although the current adaptive optical image post-processing method has advanced to a certain extent, a wide expansion space exists in the conventional algorithm, the deep learning algorithm and the fusion direction between the conventional algorithm and the deep learning algorithm. At present, the deep learning algorithm of AO image recovery is mainly supervised learning, and the theoretical requirement for training a large amount of real data is difficult to meet. In view of the complex tasks of the adaptive optics system, such as observation of remote celestial bodies, the acquisition