Lightweight bidirectional feedback network for image super-resolution. (September 2022)

• Record Type: Journal Article
• Authors: Wang, Beibei; Yan, Binyu; Liu, Changjun; Hwangbo, Ryul; Jeon, Gwanggil; Yang, Xiaomin
• Is Part Of: Computers & electrical engineering. Volume 102(2022)
• Publication Date: 2022-09
• Subjects: 00-01 -- 99-00; Super-resolution -- Cross-learning -- Regression loss -- Bidirectional feedback -- Lightweight; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2022.108254
• Languages: English
• ISSNs: 0045-7906

Abstract

Abstract: Super-resolution has attracted academic attention recently, for its capabilities of image restoration and image enhancement. To generate informative high-level features for a better reconstruction performance, most super-resolution networks have many parameters, which limit their application in resource-constrained devices. Feedback networks can generate informative high-level features with few parameters by feeding high-level features back to previous layers. In this paper, we propose a lightweight bidirectional feedback network for image super-resolution (LBFN), which consists of two feedback procedures connected in reverse. Bidirectional feedback architecture further improves the perceptual abilities of feedback networks by fusing different level features sufficiently. We propose a residual attention block to enhance the detail expression ability of feature maps, which are cross-learned in feedback block. Finally, we propose a SR regression loss to supervise the training of our network. Extensive experiments demonstrate that our method has an outstanding performance while taking up little computing resources. Graphical abstract: Highlights: Bidirectional feedback can improve the perceptual abilities of feedback networks. Cross-learning between HR and LR flow can achieve a powerful HR representation. Residual blocks with channel and spatial attention can improve the detail expression ability of features. The mapping function from LR to HR can be better learned with the help of SR regression loss.