Research On The Key Technologies Of Dim Space Target Detection Based On Deep Learning

With the rapid development of space technology in various countries,a large number of space vehicles have been sent into space,and the amount of space debris caused by satellite collision and disintegration has increased significantly,which seriously affects the normal on-orbit operation of space flight equipment.Monitoring space targets is a prerequisite to avoid satellite collision in space and an important link to maintain space environment security.As an important equipment for space target monitoring,ground-based telescope has the advantages of long detection distance,low cost and high sensitivity.However,due to the working conditions and imaging characteristics of the telescope,there are some challenges in the practical engineering.Firstly,the telescope is easily affected by optical vignetting and stray light inside and outside the field of view during imaging,which leads to the uneven distribution of the background gray,which brings difficulties for the subsequent image processing and target detection.On the other hand,because the space target of high orbit is far from the ground and the energy is weak,the image is only points or lines in the image.When the image quality is poor,it is often submerged in the noise and difficult to be accurately detected.This paper carries out in-depth analysis and research on the key technologies of space faint target detection,and focuses on solving problems in practical engineering applications.The main research work is as follows:1.Aiming at the problem of non-uniformity in the imaging process of groundbased telescopes with large field of view,this paper proposes a method of image nonuniform correction based on conditional generative adversarial net.Firstly,a heterogeneous physical model is designed and a data set for supervised learning is created.In addition,a new generator network is proposed to realize the mapping from heterogeneous image to heterogeneous background.By learning the features of the non-uniform image,the image can be corrected without obtaining the parameters of the optical system,and the image can be corrected under any complex vignetting and scattering light coupling conditions.Simulation and experimental results show that this method can suppress the heterogeneous background quickly and effectively.After correction,the mean and standard deviation of the non-uniform image obtained by ground-based telescope are 0.26 and 0.92,respectively,and the target signal-to-noise ratio is increased by 43.8%compared with the uncorrected image.2.Aiming at the problem of segmentation and detection of high-orbit dim and weak objects in the star mode of ground-based telescopes,a detection algorithm based on codec-convolutional neural network based on single frame information is proposed.Firstly,the traditional U-Net network is improved by removing the convolution layer with the smallest feature map size and introducing attention mechanism at the jump junction to enhance the utilization of shallow features.Furthermore,two methods of connecting domain labeling and end to end detection head output target position information are proposed.The improved method can not only calculate coordinates faster,but also alleviate the problem of missing detection caused by target adhesion to a certain extent.The proposed method realizes the effective segmentation and detection of dim and weak objects in a single frame image by learning the features of dim and weak objects in the image.Experimental results show that the segmentation performance of this method is better than that of some existing segmentation algorithms,and it can quickly segment bar space objects with very low signal-to-noise ratio.The accuracy of 98.5% and the false alarm of 1.6% were obtained in the actual images obtained by the telescope3.Aiming at the detection problem of high-orbit dim and weak targets in the gaze mode of ground-based telescopes,a point target detection method based on multi-frame information combined convolutional neural network and long and short term memory network is proposed in this paper.After the feature map is output by the improved YOLO-v5 s,it is sent into the AC-LSTM network for space target coordinates and confidence correlation.The fusion of time sequence information can effectively correct the problems of low confidence and false noise in static target detection.Experimental results show that this method can effectively detect point targets in gaze mode.In Spot GEO open data set,the proposed method improved the accuracy of target detection by 10% compared with the single-frame method,and obtained 97.5% detection rate and 3.2% false alarm detection results in the data obtained by the large field ground-based telescope.