Deep-Dark-Net:一种基于生成对抗网络的导星相机暗流预测模型  

Deep-Dark-Net:A Survey Camera Dark Noise Prediction Model Based on Generative Adversarial Networks

作　　者：曲伯桓 杨贺珺 何宇轩 郭远昊 刘宇[1] 曹子皇[3,5] 齐朝祥 于涌[4,5] 王培培 赵永恒[3,5] 张勇 王淑青 栗剑[3] 吕冠儒 曹兴华 向铭[3] 邱虹云 QU Bohuan;YANG Hejun;HE Yuxuan;GUO Yuanhao;LIU Yu;CAO Zihuang;QI Zhaoxiang;YU Yong;WANG Peipei;ZHAO Yongheng;ZHANG Yong;WANG Shuqing;LI Jian;LÜ Guanru;CAO Xinghua;XIANG Ming;QIU Hongyun(International School of Information Science and Engineering,Dalian University of Technology,Dalian 116620,China;Beijing University of Technology,Beijing 100124,China;National Astro-nomical Observatories,Chinese Academy of Sciences,Beijing 100191,China;Shanghai Astronomical Observatory,Chinese Academy of Sciences,Shanghai 200030,China;University of Chinese Academy of Sciences,Beijing 100049,China;Light Speed Vision(Beijing)Co.,Ltd.,Beijing 102206,China)

机构地区：[1]大连理工大学国际信息与软件学院,大连116620 [2]北京工业大学,北京100124 [3]中国科学院国家天文台,北京100191 [4]中国科学院上海天文台,上海200030 [5]中国科学院大学,北京100049 [6]光速视觉(北京)科技有限公司,北京102206

出　　处：《天文学进展》2024年第4期683-697,共15页Progress In Astronomy

基　　金：国家自然科学基金(12073047)。

摘　　要：暗流会影响图像质量、降低星像的信噪比,进而影响星像位置和流量测量的精度,因此需要在天文数据处理中准确估计并去除暗流。LAMOST导星图像处理的需求为:在无暗场图像情况下高精度处理历史导星图像数据,简化导星相机暗场图像拍摄的步骤,可以利用导星图像的特性反演和生成高精度可靠的暗场图像。利用LAMOST导星原始数据的特性,提出一种基于生成对抗网络模型来精确估计暗流的新方法——Deep-Dark-Net。该方法利用条件生成对抗网络,构建导星图像Overscan区域、Optical Black区域与对应的有效成像区域噪声之间的关联模型,从而通过这些区域反演和重构高精度暗场图像。实验表明:Deep-Dark-Net预测的暗流与真实暗流的符合度高于传统方法,满足了LAMOST望远镜导星图像处理对暗场图像的需求。该工作不仅为天文图像暗流的处理提供了一种新思路、新方法,也为深度学习技术在天文图像处理中的潜在价值和应用方向提供了重要的视角和示例。The dark current affects image quality,reduces the signal-to-noise ratio of star images,and impacts the accuracy of star position and flux measurements.Therefore,estimating and eliminating dark currents accurately is crucial when processing astronomical data.To meet the demands of LAMOST guide star image processing,this paper proposes a novel method for processing historical guide star images with high precision in the absence of dark field images,simplifying the capture process of dark field images using guide star cameras.Utilizing the characteristics of LAMOST guide star raw data,we introduce a new approach based on a generative adversarial network model,Deep-Dark-Net,to precisely estimate dark current.This method employs a conditional generative adversarial network to construct a correlation model between the Overscan and Optical Black areas of guide star images and their corresponding imaging areas.This model allows inversion and reconstruction of high-precision dark field images through these areas.Experiments demonstrate that the dark current predicted by Deep-Dark-Net aligns more closely with actual dark current than traditional methods,fulfilling the requirements for dark field image processing in LAMOST telescope guide star imagery.This work offers a new approach and methodology for handling astronomical image dark current.It highlights deep learning technology’s potential value and application direction in astronomical image processing.

关 键 词：暗流 深度学习 条件生成对抗网络 Deep-Dark-Net LAMOST 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论] P111.21[自动化与计算机技术—计算机科学与技术]