Efficient Swin Transformer for Remote Sensing Image Super-Resolution.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Super-resolution (SR) techniques are crucial for enhancing spatial details in remote sensing images.
• Transformer models show promise in remote sensing SR but often incur high computational costs.
• Existing Transformer-based methods face challenges with computational burden and memory consumption.

Purpose of the Study:

• To propose an efficient Swin Transformer network (ESTNet) for remote sensing image super-resolution.
• To reduce the computational burden and memory footprint of Transformer-based SR methods.
• To improve the quality of high-resolution remote sensing images generated from low-resolution inputs.

Main Methods:

• A three-layer convolutional operation extracts shallow image features.
• A residual group-wise attention module, incorporating efficient channel attention (ECAB) and group-wise attention (GAB), extracts deep features.
• Deep features are reconstructed to generate super-resolved remote sensing images.

Main Results:

• ESTNet achieves superior super-resolution results with a significantly lower computational burden.
• The proposed method reduces parameters by 82.68% and computational cost by 87.84% compared to existing Transformer-based SR methods.
• Experimental results validate the effectiveness and efficiency of ESTNet.

Conclusions:

• ESTNet offers an efficient and effective solution for remote sensing image super-resolution.
• The novel architecture significantly alleviates the computational challenges associated with Transformer models in this domain.
• The proposed method provides a practical approach for generating high-resolution remote sensing data.