Research On Neural Network Blind Convolution Restoration Of Atmospheric Turbulence Blurred Images

In recent years,although astronomical imaging,space target identification and earth observation imaging technologies have made unprecedented developments,the disturbance effects of atmospheric turbulence on optical systems have not been completely resolved.As we all know,the atmosphere must be considered as the transmission path of any optical system.The propagation path and direction of the light will change with the random movement of the turbulent medium,causing the image to have geometric distortion,spatial and time-varying defocus blur,and motion blur.Therefore,removing the effects of atmospheric turbulence as a wavefront disturbance is very meaningful for improving the performance of any remote imaging system(such as ground-based and space-based).The study found that due to the uncertainty of other influencing factors such as atmospheric turbulence,it is difficult for traditional image restoration methods to achieve ideal restoration results.In recent years,due to the powerful feature learning capabilities of deep neural networks,which have achieved great success not only in advanced visual tasks,but also in low-level visual tasks such as image restoration.Inspired by this,we have done a lot of research on how to use deep neural networks to reconstruct atmospheric turbulence degradation images.The main work and contributions are summarized as follows:(1)The basic theoretical knowledge of image degradation and restoration is overviewed and summarized.This paper first systematically analyzes the image degradation model,image noise model,and PSF prior model,and then determines the atmospheric turbulence degradation model of the simulated turbulence degradation image.Then,we analyzed the principles and characteristics of existing image blind restoration algorithms.Finally,through the analysis of the evaluation index of image restoration quality,the evaluation index adopted in this paper is determined.(2)A novel end-to-end convolutional neural network is designed using noise suppression blocks and U-net system to reconstruct turbulence degradation images.The network is mainly composed of FENSB,Asymmetric U-net and IRSubnetwork.In addition to the effects of atmospheric turbulence during the imaging of space targets,they are also susceptible to various types of noise.To reduce noise interference during image reconstruction,two noise suppression blocks,FENSB(feature extraction noise suppression block)and DNSB(denoising suppression block),are specially designed.The comparative experimental analysis of comprehensive simulation and real test data shows that this method has strong anti-noise and image reconstruction capabilities.(3)Aiming at the problem that the target image is seriously degraded and difficult to reconstruct due to various influencing factors,this paper proposes a simple to complex training method.This training method enables the trained network to reconstruct severely degraded images.Experiments have found that pre-training can not only make the network have better generalization performance,but also obtain higher quality reconstructed images.(4)The complex task decomposition regularization optimization strategy(TDROS)is used to reconstruct the degraded turbulent image.Inspired by the multi-task regularization idea,we propose a general TDROS,which can decompose complex image restoration problems into simple sub-problems to solve.The restoration results of turbulence degraded images by adding 11-layer and 22-layer CNNs of TDROS fully prove the feasibility and superiority of TDROS.At the same time,we use TDROS to design a novel deep neural network model to reconstruct turbulent degraded images.To avoid over-processing and under-processing of potential tasks,we use weighted methods to constrain potential tasks.Compared with advanced image restoration algorithms,our method achieves better results.This method is more competitive especially in the case of severe turbulence degradation.