RGSR: A two-step lossy JPG image super-resolution based on noise reduction

作     者：Li, Biao Shi, Yong Wang, Bo Qi, Zhiquan Liu, Jiabin 

作者机构：Southwestern Univ Finance & Econ Sch Business Adm Chengdu 611130 Peoples R China Univ Chinese Acad Sci Sch Econ & Management Beijing 101408 Peoples R China Chinese Acad Sci Res Ctr Fictitious Econ & Data Sci Beijing 100190 Peoples R China Chinese Acad Sci Key Lab Big Data Min & Knowledge Management Beijing 100190 Peoples R China Univ Nebraska Coll Informat Sci & Technol Omaha NE 68182 USA Univ Int Business & Econ Sch Informat Technol & Management Beijing 100029 Peoples R China Samsung Res China Beijing SRC B Beijing 100028 Peoples R China Southwest Minzu Univ Coll Elect & Informat Engn Chengdu 610041 Peoples R China 

出 版 物：《NEUROCOMPUTING》 (神经计算)

年 卷 期：2021年第419卷

页      面：322-334页

核心收录：

学科分类：08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Natural Science Foundation of China [71932008, 71801232, 61702099, 71501175, 91546201] Open Project of Key Laboratory of Big Data Mining and Knowledge Management, Chinese Academy of Sciences Fundamental Research Funds for the Central Universities in UIBE [CXTD10-05] 

主　　题：Super-resolution JPEG compression Cycle loss Image denoising 

摘      要：Single Image Super-Resolution (SISR) is a fundamental and important low-level computer vision (CV) task, yet its performance on real-world applications is not always satisfactory. Different from the previous SISR research, we focus on a specific but realistic SR issue: How can we obtain satisfied SR results from com-pressed JPG (C-JPG) images, which is a ubiquitous image format to greatly release storage space while missing fine details. the JPG SR task is deeply analyzed to discover the connotation. Then, we propose an effective two-step model structure named RGSR, involving two specifically designed components, i.e., JPG recovering and SR generation, instead of the perspective of noise elimination in traditional SR approaches. Besides, we further integrate the cycle loss to build a hybrid objective across scales for better SR generation. Experimental results on both of the standard test data sets and real images show that our approach achieves outstanding results and succeed in applying to practical C-JPG SR tasks. (c) 2020 Elsevier B.V. All rights reserved.