Open AccessImproved Image Super-Resolution Using Frequency Channel Attention and Residual Dense Networks
Author(s) -
Zhikuan Sun,
Zheng Li,
Mengchuan Sun,
Ziwei Hu
Publication year - 2022
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/2216/1/012074
Subject(s) - residual , computer science , artificial intelligence , image (mathematics) , channel (broadcasting) , focus (optics) , computer vision , image quality , convergence (economics) , resolution (logic) , image resolution , low resolution , pattern recognition (psychology) , high resolution , algorithm , telecommunications , remote sensing , optics , physics , geology , economics , economic growth
In our real life, due to the various influences, low-resolution images exist widely. The resolution of the image represents the amount of information carried by the image and the quality of the image. Image super-resolution means reconstructing low-resolution images, and it can help us improve the image quality to get more information. This paper attempts to improve the existing super-resolution reconstruction model based on deep learning. We use nested residual dense connection to prompt the model to focus on the recovery of detailed textures and accelerate convergence. Meanwhile, we use frequency channel attention mechanism to weight channels. By comparing the experimental results with other methods including FSRCNN, VDSR and MemNet, our proposed method has achieved better results and visual improvements. Especially on the Urban100 test dataset, the increases of PSNR and SSIM reach higher.