A Remote Sensing Image Super-Resolution Reconstruction Model Combining Multiple Attention Mechanisms.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Remote sensing images present challenges for super-resolution due to complexity and scale variations.
• Existing deep learning methods struggle with global context and detail utilization.

Purpose of the Study:

• To develop a novel super-resolution reconstruction model for remote sensing data.
• To address limitations in existing deep learning-based super-resolution algorithms.

Main Methods:

• Introduced a multi-branch residual hybrid attention block (MBRHAB).
• Employed window-based multi-head self-attention for long-range dependencies.
• Utilized a multi-branch convolution module (MBCM) to enhance receptive fields.
• Combined channel and spatial attention for augmented local semantic information.
• Implemented a parallel design for computational efficiency.

Main Results:

• The proposed method demonstrated superior performance over existing algorithms (Bicubic, SRCNN, ESRGAN, Real-ESRGAN, IRN, DSSR).
• Achieved higher PSNR and SSIM metrics across various magnification scales.
• Validated generalization performance using cross-dataset experiments.

Conclusions:

• The MBRHAB model effectively enhances super-resolution reconstruction for remote sensing images.
• The proposed approach offers improved utilization of global and local details.
• The method shows strong generalization capabilities for diverse remote sensing datasets.