Cross-domain continuous-scale remote sensing image super-resolution via meta-weight learning.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Continuous-scale Remote Sensing Image Super-Resolution (RSISR) offers flexibility but struggles with spatial anchor feature aggregation.
• Existing methods use fixed weighting, limiting detail recovery and failing in cross-domain applications.

Purpose of the Study:

• To develop a novel meta-learning approach for adaptive feature aggregation in RSISR.
• To improve spatial adaptability and bridge the cross-domain gap without encoder fine-tuning.

Main Methods:

• Introduced Meta-Weight Learning, a meta-learning framework.
• Developed a novel implicit decoder to dynamically predict content-aware weights.
• Weights integrate feature representations and relative geometric positions for adaptive fusion.

Main Results:

• Achieved state-of-the-art performance in continuous-scale RSISR.
• Demonstrated superior spatial adaptability and effective cross-domain transfer.
• Outperformed existing methods in both visual quality and quantitative metrics.

Conclusions:

• Meta-Weight Learning effectively addresses limitations of fixed weighting in RSISR.
• The proposed framework significantly enhances detail recovery and cross-domain performance.
• This approach offers a robust solution for continuous-scale RSISR across diverse domains.