TY - JOUR
T1 - A Lightweight Pyramid Feature Fusion Network for Single Image Super-Resolution Reconstruction
AU - Liu, Bingzan
AU - Ning, Xin
AU - Ma, Shichao
AU - Lian, Xiaobin
N1 - Publisher Copyright:
© 1994-2012 IEEE.
PY - 2024
Y1 - 2024
N2 - With the development of deep learning and super-resolution reconstruction, the performance of single-image super-resolution (SISR) has improved significantly. However, most of them cannot strike a great balance between computational cost and performance, which prevents them from being deployed on edge devices. Additionally, feature fusion and channel mixing are not considered in most lightweight networks, leading to the limited performance of these networks. To solve such problems, we propose a lightweight pyramid feature fusion network (PFFN), which mainly contains the pyramid spatial-adaptive feature extraction module (PSAFEM) and the enhanced channel fusion module (ECFM). They can extract global-to-local feature, build long-range dependence with small parameters increment and realize channel and spatial feature fusion. Finally, some state-of-the-art methods are utilized to compare with our network. Extensive experimental results indicate that our PFFN outperforms these methods in parameters, flops and performance.
AB - With the development of deep learning and super-resolution reconstruction, the performance of single-image super-resolution (SISR) has improved significantly. However, most of them cannot strike a great balance between computational cost and performance, which prevents them from being deployed on edge devices. Additionally, feature fusion and channel mixing are not considered in most lightweight networks, leading to the limited performance of these networks. To solve such problems, we propose a lightweight pyramid feature fusion network (PFFN), which mainly contains the pyramid spatial-adaptive feature extraction module (PSAFEM) and the enhanced channel fusion module (ECFM). They can extract global-to-local feature, build long-range dependence with small parameters increment and realize channel and spatial feature fusion. Finally, some state-of-the-art methods are utilized to compare with our network. Extensive experimental results indicate that our PFFN outperforms these methods in parameters, flops and performance.
KW - Lightweight super-resolution reconstruction
KW - multi-scale feature fusion
KW - pyramid feature extraction
UR - http://www.scopus.com/inward/record.url?scp=85195414453&partnerID=8YFLogxK
U2 - 10.1109/LSP.2024.3410017
DO - 10.1109/LSP.2024.3410017
M3 - 文章
AN - SCOPUS:85195414453
SN - 1070-9908
VL - 31
SP - 1575
EP - 1579
JO - IEEE Signal Processing Letters
JF - IEEE Signal Processing Letters
ER -