TY - GEN
T1 - Cascade Scale-Aware Distillation Network for Lightweight Remote Sensing Image Super-Resolution
AU - Ji, Haowei
AU - Di, Huijun
AU - Wang, Shunzhou
AU - Shi, Qingxuan
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022.
PY - 2022
Y1 - 2022
N2 - Recently, convolution neural network based methods have dominated the remote sensing image super-resolution (RSISR). However, most of them own complex network structures and a large number of network parameters, which is not friendly to computational resources limited scenarios. Besides, scale variations of objects in the remote sensing image are still challenging for most methods to generate high-quality super-resolution results. To this end, we propose a scale-aware group convolution (SGC) for RSISR. Specifically, each SGC firstly uses group convolutions with different dilation rates for extracting multi-scale features. Then, a scale-aware feature guidance approach and enhancement approach are leveraged to enhance the representation ability of different scale features. Based on SGC, a cascaded scale-aware distillation network (CSDN) is designed, which is composed of multiple SGC based cascade scale-aware distillation blocks (CSDBs). The output of each CSDB will be fused via the backward feature fusion module for final image super-resolution reconstruction. Extensive experiments are performed on the commonly-used UC Merced dataset. Quantitative and qualitative experiment results demonstrate the effectiveness of our method.
AB - Recently, convolution neural network based methods have dominated the remote sensing image super-resolution (RSISR). However, most of them own complex network structures and a large number of network parameters, which is not friendly to computational resources limited scenarios. Besides, scale variations of objects in the remote sensing image are still challenging for most methods to generate high-quality super-resolution results. To this end, we propose a scale-aware group convolution (SGC) for RSISR. Specifically, each SGC firstly uses group convolutions with different dilation rates for extracting multi-scale features. Then, a scale-aware feature guidance approach and enhancement approach are leveraged to enhance the representation ability of different scale features. Based on SGC, a cascaded scale-aware distillation network (CSDN) is designed, which is composed of multiple SGC based cascade scale-aware distillation blocks (CSDBs). The output of each CSDB will be fused via the backward feature fusion module for final image super-resolution reconstruction. Extensive experiments are performed on the commonly-used UC Merced dataset. Quantitative and qualitative experiment results demonstrate the effectiveness of our method.
KW - Feature distillation
KW - Lightweight neural network
KW - Multi-scale feature learning
KW - Remote sensing image super-resolution
UR - http://www.scopus.com/pages/publications/85142846790
U2 - 10.1007/978-3-031-18916-6_23
DO - 10.1007/978-3-031-18916-6_23
M3 - Conference contribution
AN - SCOPUS:85142846790
SN - 9783031189159
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 274
EP - 286
BT - Pattern Recognition and Computer Vision - 5th Chinese Conference, PRCV 2022, Proceedings
A2 - Yu, Shiqi
A2 - Zhang, Jianguo
A2 - Zhang, Zhaoxiang
A2 - Tan, Tieniu
A2 - Yuen, Pong C.
A2 - Guo, Yike
A2 - Han, Junwei
A2 - Lai, Jianhuang
PB - Springer Science and Business Media Deutschland GmbH
T2 - 5th Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2022
Y2 - 4 November 2022 through 7 November 2022
ER -