基于改进EDSR的轻量化机载图像超分辨率算法  

A Lightweight EDSR-Based Algorithm for Super-Resolution Reconstruction of Airborne Image

作　　者：王辰北 张海军[1] 王浩然 WANG Chenbei;ZHANG Haijun;WANG Haoran(Luoyang Institute of Electro-Optical Equipment AVIC,Luoyang 471000 China;Science and Technology on Electro-Optical Control Laboratory,Luoyang 471000 China)

机构地区：[1]中国航空工业集团公司洛阳电光设备研究所,河南洛阳471000 [2]光电控制技术重点实验室,河南洛阳471000

出　　处：《电光与控制》2022年第10期39-43,70,共6页Electronics Optics & Control

摘　　要：近年来,高分辨率数字化显示器广泛应用于机载座舱中,受限于特定波段图像传感器的性能以及视频传输链路等因素,到达显示器端的机载传感器图像的分辨率往往低于显示器分辨率。这些图像直接显示时有效显示区域较小,不方便观察;简单的插值算法放大又会导致画面模糊;而超分辨率(SR)重建技术在放大图像的同时,预测了图像细节,从而提升图像的分辨率。目前,基于卷积神经网络(CNN)的超分辨率算法凭借其卓越的重建效果(PSNR及SSIM)成为最先进的SR算法;现有的SR算法存在网络结构复杂、参数量大、计算消耗资源大等问题,难以在机载嵌入式环境下实时实现。针对以上问题,提出一种轻量化超分辨率算法C-EDSR,在重建效果基本不降低的情况下,大幅减少计算量,为进一步在机载嵌入式环境下实时实现SR算法奠定了基础。实验表明:所提算法与EDSR相比,PSNR值平均降低0.026 dB,SSIM值平均降低0.000176,计算量减少61.1%。In recent years,high-resolution digital displays have been widely used in airborne cockpit.Limited by the performance of the image sensor in a specific band and the video transmission link the resolution of the airborne sensor image arriving at the display is often lower than that of the display.When these images are directly displayed the effective display area is small which is inconvenient for observation and simply using the interpolation algorithm for image magnification will cause image blurring.Super Resolution(SR)reconstruction technology can predict image details while magnifying the image at the same time which effectively improves image resolution.Currently the CNN-based SR algorithm has become the most advanced SR algorithm owing to its excellent reconstruction effects(PSNR&SSIM).The existing SR algorithm has the problems of complex network structure a large amount of parameters and huge computational resource consumption which is inapplicable for real-time implementation in airborne embedded environment.To solve the above problems a lightweight super-resolution algorithm named C-EDSR is proposed which greatly reduces the amount of computation without harming reconstruction effects.The new algorithm provides a basis for further real-time implementation of SR algorithm in airborne embedded environment.The experimental results show that the proposed C-EDSR reduces 61.1%computation while decreasing PSNR by 0.026 dB and SSIM by 0.000176 on average compared with EDSR.

关 键 词：超分辨率 卷积神经网络 轻量化 机载环境 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]