Study On The Co-phasing Error Detection Technology Of Synthetic Aperture Telescopes Based On Deep Learning

With the development of astronomical observation and remote sensing monitoring,it plays an increasingly important role in modern technology.The requirements for the resolution performance of optical telescopes are further improved,and the development of telescopes has shown a trend of increasing aperture size to improve the imaging resolution.Due to the restrictions on factors as manufacturing technology,system assembly and spacecraft payload and some others,it is difficult for traditional singleaperture telescopes to achieve further breakthroughs in expanding the aperture.Synthetic aperture imaging technology is able of achieving equivalent high-resolution imaging performance of large-aperture telescope systems with multiple separated subaperture arrays,and will surely become one of the main development directions of highresolution imaging in the future.However,the existence of the co-phasing error in the synthetic aperture system will be a big barrier to the improvement in imaging resolution.Only by eliminating the influence of co-phasing error can the synthetic aperture system truly achieve high-resolution imaging.Centering on the co-phasing error detection,which is one of the key technologies in synthetic aperture system technology,we introduce deep learning method into the co-phasing error detection,and implement some researches by means of simulations and experiments.The main contents are as follows:First,the research progress of optical synthetic aperture imaging system has been reviewed,and the characteristics and limitations of the existing traditional co-phasing detection technologies as well as co-phasing detection method using neural network are s?mmarized.According to the development tendency of the co-phasing error detection and the major challenges it faces,a new approach using deep learning is proposed,so as to explore a more practical co-phasing method.Secondly,by combining the object-image relationship of the synthetic aperture system and the principle of deep learning,a mathematical physical model of optical synthetic aperture co-phasing error detection using deep learning has been established.Using a single convolutional neural network to fit the input-output mapping relationship between the broadband point spread function images and pistons of all sub-apertures,the proposed approach can realize an end-to-end co-phasing detection.This method is capable of directly achieving the fine co-phasing of the system over a large capture range,and only one frame of focal point spread function image is need for each cophasing detection,thus decreasing the complexity of the imaging system and realizing rapid detection of co-phasing.Besides,the limiting detection range beyond coherent length of the method has been studied,which further expands the detection range of the co-phasing error.Then,a two-aperture experimental platform and a three-aperture experimental platform are built to collect real images for training and testing,and it is demonstrated that the co-phasing detection technique using a single convolutional neural network is feasible to practical systems by experimental results.Besides,we analyzed the difficulties in collecting amounts of real images,and propose a solution by means of learning to sense pistons from simulation.By constructing simulation models as possible as close to the experimental setup,simulated images are generated and applied to the training of network instead of the experimental training data.The co-phasing detection results on experimental testing dataset demonstrate that the network trained by the simulation dataset can accurately sense pistons from the real images.Finally,to solve the difficulties in co-phasing detection of extended targets,focal image and defocus image are used to construct the feature image training dataset which is independent of the target characteristics.It is demonstrated that a single convolutional neural network is capable of directly extracting the co-phasing error of all sub-aperture from the broadband feature images by simulations.Besides,we analyzed the influences of the increasing complexity in the mapping relationship,which is caused by the increase in the n?mber of sub-apertures,on the detection accuracy and detection speed of the proposed technique.Some simulations have been implemented based on centrosymmetric redundant array and noncentrosymmetric redundant array to analyze the influence of baseline redundancy on the co-phasing detection performance of the proposed approach.Combining the simulated results and theoretical analysis,we propose an effective solution to this problem.