TTST: A Top-k Token Selective Transformer for Remote Sensing Image Super-Resolution.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Transformer models show promise in image super-resolution due to global aggregation.
• Challenges exist in applying Transformers to large-area remote sensing: redundant tokens and single-scale representation.

Purpose of the Study:

• To develop an efficient Transformer-based method for large-area remote sensing image super-resolution.
• To address redundant token representation and single-scale feature issues in existing Transformers.

Main Methods:

• Proposed Residual Token Selective Group (RTSG) for compact self-attention by selecting top-k keys.
• Developed Multi-scale Feed-forward Layer (MFL) for enriched multi-scale feature representation.
• Introduced Global Context Attention (GCA) to enhance informative components for accurate reconstruction.

Main Results:

• The proposed Top-k Token Selective Transformer (TTST) achieves favorable performance against state-of-the-art methods on simulated and real-world remote sensing datasets.
• TTST demonstrates improved image super-resolution quantitatively and qualitatively.
• TTST achieves 0.14 dB higher PSNR than HAT-L while using only 47.26% of its computational cost and 46.97% of its parameters.

Conclusions:

• TTST offers an effective and efficient solution for remote sensing image super-resolution.
• The method successfully mitigates challenges of redundant tokens and single-scale representation.
• TTST provides a strong baseline for future research in Transformer-based remote sensing image enhancement.