Panchromatic Image Super-Resolution Via Self Attention-Augmented Wasserstein Generative Adversarial Network.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Panchromatic (PAN) images offer valuable spatial data for earth observation but are often low-resolution (LR).
• Existing super-resolution (SR) methods struggle to perfectly reconstruct fine edge details in SR images.

Purpose of the Study:

• To develop an improved SR model for enhancing edge details in LR images.
• To leverage multi-feature relevance for superior detail reconstruction in super-resolution tasks.

Main Methods:

• An encoder-decoder network with a fully convolutional network (FCN) backbone extracts multi-scale features.
• A Convolutional Block Attention Module (CBAM) is integrated into skip-connections to enhance feature representation.
• A novel Self-Attention Augmented (SAA) module dynamically generates attention weights based on feature similarity for improved detail preservation.

Main Results:

• The proposed SAA-WGAN model effectively extracts and utilizes multi-layer feature relevance.
• The method demonstrates superior performance in reconstructing edge details compared to existing SR techniques.
• Experimental results show significant improvements in both objective metrics and visual quality.

Conclusions:

• The SAA-WGAN model offers a robust solution for enhancing spatial information in LR earth observation images.
• The integration of CBAM and the novel SAA module significantly boosts the detail reconstruction capabilities of SR models.
• This approach advances the field of super-resolution for remote sensing applications.