图像背景在图像超分辨率中的作用研究  

Investigating the Function of Image Background in Image Super-Resolution

作　　者：刘茜娜 顾津锦 董超 LIU Xina;GU Jinjin;DONG Chao(Shenzhen Institute of Advanced Technology,Chinese Academy of Sciences,Shenzhen 518055,China;University of Chinese Academy of Sciences,Beijing 100049,China;The University of Sydney,Sydney 2006,Australia)

机构地区：[1]中国科学院深圳先进技术研究院,深圳518055 [2]中国科学院大学,北京100049 [3]悉尼大学,悉尼2006

出　　处：《集成技术》2023年第5期76-91,共16页Journal of Integration Technology

摘　　要：图像超分辨率是底层视觉领域的一项代表性任务,相关研究发现,图像某个像素位置的重建质量与其周围的背景有关。基于这一发现,该文探索了通过分割输入图像解释网络的新视角,提出了一种简单组合数据集,该数据集信息量丰富,但单张图中仅包含单一的纹理信息。实验证明,与目标区域纹理相近的背景,较有利于模型在该区域的超分辨率重建;对比分析注意力机制与传统卷积神经网络,结果显示,注意力结构更能帮助网络关注长程有效信息。As a representative low-level vision problem,image super-resolution(SR)aims to reconstruct the high-resolution image from its low-resolution counterpart.For a long time,the analysis of SR tasks is based on the whole image,while little works observe the input partition.In this paper,we find that the restoration quality of a certain position is inseparable from its surrounding image background.This phenomenon provides us a new perspective to explain the networks by spliting the input image.We construct a new hybrid dataset,of which the foreground and background contain only one kind of texture information.And then,we prove that the similar background could benefit the network restoration.By analyzing similarity and difference between the attention mechanism and the traditional CNN network,we show that the attention structure could help the network focus on long-range effective information.Moreover,a data enhancement method to improve the network final performance and potential future works are also proposed.

关 键 词：超分辨率 网络可解释性 图像背景 数据集 

分 类 号：TP39[自动化与计算机技术—计算机应用技术] TP751.1[自动化与计算机技术—计算机科学与技术]