HSFAN: A Dual-Branch Hybrid-Scale Feature Aggregation Network for Remote Sensing Image Super-Resolution.

Area of Science:

• Computer Vision
• Remote Sensing
• Image Processing

Background:

• Existing remote sensing image super-resolution networks struggle with high-entropy regions, detail recovery, and computational complexity.
• Insufficient feature utilization and excessive model parameters limit current methods.

Purpose of the Study:

• To propose a novel dual-branch hybrid-scale feature aggregation network (HSFAN) for remote sensing image super-resolution.
• To achieve an optimal balance between model complexity and reconstruction quality.
• To enhance feature utilization in high-entropy regions and improve detail recovery.

Main Methods:

• The HSFAN employs a dual-branch architecture: a main branch with a multi-scale parallel large convolution kernel (MSPLCK) module for global structure modeling and an enhanced parallel attention (EPA) module for feature prioritization.
• An auxiliary branch utilizes a multi-scale large-kernel attention (MSLA) module with depthwise separable convolutions to reduce computational overhead and capture local high-frequency information.
• Consistency constraints are maintained in the feature space throughout the main branch.

Main Results:

• The proposed HSFAN achieved a PSNR of 27.91 dB and an SSIM of 0.7616 on the UC Merced dataset for ×4 super-resolution.
• The algorithm outperformed most current mainstream super-resolution algorithms.
• HSFAN maintained a low computational cost and model parameter count compared to existing methods.

Conclusions:

• The HSFAN effectively addresses limitations in existing remote sensing image super-resolution networks.
• The proposed network offers a new technical route for improved remote sensing image super-resolution reconstruction.
• HSFAN provides a promising balance between reconstruction quality and computational efficiency.