A lightweight high-frequency mamba network for image super-resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Single Image Super-Resolution (SISR) research focuses on enhancing image resolution.
• Existing methods often struggle to effectively combine local and global image information.
• Convolutional Neural Networks (CNNs) and Transformers are common but have limitations in integrating diverse information types.

Purpose of the Study:

• To develop a novel approach for SISR that optimally integrates local and global image features.
• To address the computational complexity associated with Transformer-based models in SISR.
• To propose a lightweight yet effective network for high-frequency detail restoration in super-resolution.

Main Methods:

• Proposed the High-frequency Mamba Network (HFMN) for SISR.
• Utilized a self-attention mechanism to balance local and global information weights.
• Employed the selective state space model VMamba for efficient global feature extraction, reducing computational complexity.
• Introduced specialized modules: Local High-Frequency Feature Block (LHFB), Mamba-Based Attention Block (MAB), and Dual-information Interactive Attention Block (DIAB).

Main Results:

• The HFMN demonstrated superior performance on benchmark SISR datasets compared to recent state-of-the-art (SOTA) methods.
• The proposed network achieved high-frequency detail restoration effectively.
• HFMN achieved these results with significantly fewer parameters, indicating a lightweight architecture.

Conclusions:

• The HFMN effectively integrates local and global information for SISR using VMamba and attention mechanisms.
• The network offers a computationally efficient and lightweight solution for super-resolution tasks.
• HFMN represents a significant advancement in SISR, outperforming existing methods with improved efficiency.