ACT-SR: Aggregation Connection Transformer for Remote Sensing Image Super-Resolution

Recently, Transformer-based methods have shown impressive performances in remote sensing image superresolution (RSISR). However, the application of Transformer in RSISR frequently results in artifacts and the loss of image detail due to limited information transmission pathways and the constraints o...

Full description

Saved in:
Bibliographic Details
Published in:IEEE journal of selected topics in applied earth observations and remote sensing 2024-11, p.1-12
Main Authors: Kang, Yingdong, Wang, Xinyu, Zhang, Xuemin, Wang, Shaoju, Jin, Guang
Format: Article
Language:English
Subjects:
Convolution
Deep learning
Feature extraction
Image reconstruction
Image restoration
Interpolation
Logic gates
optical remote sensing
Remote sensing
Super-resolution
Transformer
Transformers
Visualization
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Recently, Transformer-based methods have shown impressive performances in remote sensing image superresolution (RSISR). However, the application of Transformer in RSISR frequently results in artifacts and the loss of image detail due to limited information transmission pathways and the constraints of uni-dimensional self-attention mechanisms. To solve these problems, an Aggregation Connection Transformer (ACT-SR) is proposed for RSISR. ACT-SR employs an advanced attention mechanism designed to enrich information transmission across spatial and channel dimensions, thereby enlarging the receptive fields for more accurate feature extraction. A core component of ACT-SR is the novel aggregation connection attention block, which effectively captures spatial similarities and channel importance, aggregating this information through a combination of series and parallel connections for enhanced feature representation. Furthermore, a new gated feed-forward network is introduced to enhance the nonlinear mapping capabilities of the Transformer and control the information flow through the network. In addition, ACT-SR integrates a shifted windows scheme alongside interpolation residual calculation to facilitate efficient detail recovery and artifact elimination. Experimental results confirm the effectiveness of the proposed modules, with ACT-SR outperforming several state-of-the-art RSISR methods in both objective metrics and visual quality
ISSN:1939-1404
DOI:10.1109/JSTARS.2024.3506717