Lightweight interactive feature inference network for single-image super-resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) and Transformers have advanced image super-resolution (SR).
• Existing complex models incur high computational costs and large parameter counts.
• Current methods often neglect crucial structural priors for high-quality reconstruction.

Purpose of the Study:

• To develop a lightweight and effective image super-resolution network.
• To improve reconstruction performance by integrating CNN and Transformer architectures.
• To address the limitations of existing models in handling structural priors and computational efficiency.

Main Methods:

• A novel lightweight Interactive Feature Inference Network (IFIN) is proposed.
• The network backbone, Interactive Feature Aggregation Module (IFAM), utilizes Structure-Aware Attention Blocks (SAAB), Swin Transformer Blocks (SWTB), and Enhanced Spatial Adaptive Blocks (ESAB).
• SAAB recalibrates local structures, SWTB captures global information, and ESAB fuses diverse features.

Main Results:

• The proposed IFIN achieves state-of-the-art reconstruction accuracy on benchmark datasets.
• The network demonstrates significantly lower computational demands compared to existing methods.
• Experiments validate the effectiveness of SAAB, SWTB, and ESAB in feature extraction and fusion.

Conclusions:

• The IFIN model offers a computationally efficient solution for high-quality image super-resolution.
• The synergistic integration of CNN and Transformer components enhances the handling of structural priors.
• The developed network provides a promising direction for future research in efficient deep learning for SR.