基于望远系统的光学-神经网络联合优化超分辨成像方法  

A joint optimization super-resolution imaging method of optical-neural network based on a telescopic system

作　　者：孙友红 张涛[2] 刘嘉楠 刘建华[1] 王超[1] Sun Youhong;Zhang Tao;Liu Jianan;Liu Jianhua;Wang Chao(Changchun University of Science and Technology Institute of Space Ophotoelectronics Technology,Changchun,130022,China;The Institute of Remote Sensing Satellites,China Academy of Space Technology(CAST),Beijing,100094,China)

机构地区：[1]长春理工大学空间光电技术研究所,长春130022 [2]中国空间技术研究院(CAST)遥感卫星总体部,北京100094

出　　处：《仪器仪表学报》2024年第12期12-23,共12页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(62375027,62127813);重庆市自然科学基金(CSTB2023NSCQ-MSX0504);吉林省自然科学基金(YDZJ202201ZYTS411,222621JC010498735);吉林省教育厅资助(JJKH20240920KJ)项目资助。

摘　　要：光学望远镜是获取远距离物体光学信息的重要工具,在天文观测、遥感和光学监视领域具有广泛的应用。分辨率是衡量望远镜观测物体细节能力的重要指标,传统提高望远镜分辨率的方法是建造更大口径的望远镜,这导致建造和维护成本大幅增加。本文提出一种光学-神经网络联合优化方法,通过将望远系统的点扩散函数等效为一个单核卷积层,集成到图像超分辨重建网络前端进行联合训练,并在光路引入相位掩模重构训练得到的点扩散函数,从而实现两者协同优化,有效提高了观测图像的分辨率。本文还构建一种高性能的生成对抗网络,其训练参数小于现有几种无监督网络,重建速度远高于现有几种无监督网络。此网络采用双鉴别器架构提高了网络提取细节特征的能力,设计的级联残差块充分利用了各级提取的特征信息,扩展了信息的传播路径,提高了重建效率。仿真结果表明,本文与单纯的深度学习方法相比,联合优化方法重建的超分辨率图像PSNR和SSIM在仿真数据集中分别提高了3.98和0.06,图像细节丰富,容易分辨。验证实验表明,本文的联合优化方法重建的条纹图像对比度最高,更容易分辨。Optical telescopes are important tools for obtaining optical information about distant objects,and resolution is an important index to measure their ability to observe objects in detail.The traditional way to improve resolution is to build larger telescopes,which significantly increases construction and maintenance costs.In this paper,an optical-neural network joint optimization method is proposed.The point spread function of the telescopic system is trained jointly with the image reconstruction network,and the point spread function obtained from the phase mask reconstruction training is introduced into the optical path to realize the cooperative optimization of both and improve the resolution of the observed image.A lightweight generative adduction network is constructed,whose reconstruction speed is much higher than that of several existing unsupervised networks.LCR-GAN uses a dual discriminator architecture to improve the network′s ability to extract detailed features.The designed cascade blocks fully use the extracted feature information at all levels and improve the reconstruction efficiency.The simulation and verification results show that compared to the simple deep learning method,the super-resolution image reconstructed by this method is rich in detail,and the stars and stripes are easier to distinguish.Optical telescope is an important tool to obtain optical information of distant objects,and has wide application in astronomical observation,remote sensing,and optical surveillance.Resolution is an important indicator of a telescope's ability to observe objects in detail,and the traditional way to improve the resolution of a telescope is to build a larger aperture telescope,which leads to a significant increase in construction and maintenance costs.In this paper,an optic-neural network joint optimization method is proposed.The point diffusion function of the telescopic system is equivalent to a single-core convolution layer,which is integrated into the front end of the image super-resolution reconstruct

关 键 词：超分辨成像 联合优化 望远系统 无监督网络 

分 类 号：O439[机械工程—光学工程] TH743[理学—光学]