Atmospheric Turbulence Degradation Image Restoration Based On Machine Learning And Turbulence Parameters Inversion Of Optical Flow Algorithm

Long-distance imaging has been widely concerned by researchers at home and abroad,and the main factor causing the degradation of long-distance imaging is atmospheric turbulence.At present,the models established by traditional physical methods are too simple,resulting in unsatisfactory image restoration effects.At the same time,from the perspective of degraded image quality,the influence of turbulence on the image is often manifested through the random distortion and non-uniform blur of the image.If the space-invariant isoplanatic turbulence effect is simplified,the restoration effect will not meet the requirements.In recent years,with the extensive application of machine learning algorithms in the image field,excellent results have been achieved.Therefore,this paper successively proposes a machine learning algorithm based on supervised learning,self-supervised learning and deep transfer learning network to restore real atmospheric turbulence degradation images,and uses multiple image evaluation indicators for quantitative analysis.In addition,by combining the target tracking in computer vision,this paper proposes a Median Absolute Deviation(MAD)optical flow algorithm to invert the near-surface and highaltitude atmospheric turbulence parameters from the turbulence degradation image,respectively.The main work content and achievements of this paper are summarized as follows:1.Based on the assumption that the linear space is invariant,two neural network models are proposed.including:(Ⅰ)celestial images were taken by using the Hubble space-based telescope,and combine the long-exposure turbulence degradation model to construct a simulated datasets.Inception V2 was used by the constructed network as the backbone architecture,combined with bidirectional multi-scale feature fusion,to learn turbulence degradation features.Subsequently,the performance of the proposed network was verified by using the ISS images from the Munin ground-based telescope.The results show that compared with the CLEAR algorithm,the SGL algorithm,the IBD algorithm,and the DNCNN network,the designed network has a large lead in the restoration speed.Generally,the results can be output within 0.5s,and the average gradient and spatial frequency are also significantly improved.(Ⅱ)In order to get rid of the dependence on data-driven,a self-supervised learning network with encoder architecture is proposed,the network iteration is constrained by denoising regularization(RED)and two-stage loss,and no reference metrics(Entropy and AverageGradient)were employed for quantitative calculation,the self-supervised learning network constructed has a superior output on some images,but residual artifacts and ringing effects.2.Practically,the atmospheric turbulence for long-range imaging is mostly in the anisoplanatic region.In order to design the network as closely possible to the real turbulence degradation model,the introduced deep transfer learning was trained by using GoPro datasets combined with a small amount of Hot-Air datasets to learn the non-uniform blur and the geometric distortion,respectively.Further,four non-reference indicators are used for objective evaluation.Meanwhile,the ablation study was carried out to prove the presented network has the stronger robustness and generalization ability.3.A method for estimating atmospheric turbulence parameters based on Median Absolute Deviation(MAD)optical flow estimation is proposed.Through optical flow estimation,the pixel point drift information of the front and rear multi-frame images captured by the telescope system is captured,and the arrival time on the transmission path is obtained.Angular fluctuation variance mean,and calculate the atmospheric refractive index structure constant(Cn2)and atmospheric coherence length(r0).In the near-ground experiment,the estimated value of the algorithm is compared with the measured results of the temperature pulsation instrument.The BIAS,RMSE,and the Rxy are-0.0202,0.2391,and 0.8230,respectively.It is proved that the proposed method can effectively estimate the value and diurnal variation trend of the near-surface atmospheric refractive index structure constant(Cn2).In the upper-air observation experiment at night,the atmospheric coherence length(r0)estimated by the model was compared with the measured value of DIMM,and the integrated value of the spherical sounding,and it was found that the estimated results can roughly reflect the turbulence of the entire layer at night There is a changing trend,but the overall value is large.Subsequently,the error rationality analysis was carried out.4.The proposed optical flow algorithm is packaged into software by using the PyQt platform,which includes two estimation modes,and uses multi-threaded calculations to facilitate rapid analysis of image data.In addition,the designed software has the function of automatic networking,and the local weather conditions can be obtained before the experiment,which is convenient for the experiment plan.