REMA: A Rich Elastic Mixed Attention Module for Single Image Super-Resolution

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Detail preservation is a key challenge in single image super-resolution (SISR).
• Current deep learning SISR methods often prioritize lightweight designs, potentially compromising performance in demanding applications.
• Existing attention modules may lack the adaptability required for complex image restoration tasks.

Purpose of the Study:

• To propose a novel plug-and-play attention module, Rich Elastic Mixed Attention (REMA), for enhancing SISR performance.
• To improve detail preservation and network compatibility in super-resolution tasks.
• To investigate the relationship between module characteristics and network robustness.

Main Methods:

• Developed Rich Elastic Mixed Attention (REMA), integrating Rich Spatial Attention Module (RSAM) and Rich Channel Attention Module (RCAM).
• Introduced Rich Structure to enhance REMA's adaptability to varying input complexities and task requirements.
• Conducted extensive experiments using a REMA-based SR network (REMA-SRNet) and compared it with existing attention modules and algorithms.

Main Results:

• REMA significantly enhances performance and compatibility in super-resolution networks compared to other attention modules.
• REMA-SRNet demonstrates superior visual effects and objective evaluation quality over comparative algorithms.
• Module compatibility is found to correlate with cardinality and in-branch feature bandwidth.

Conclusions:

• REMA is an effective attention module for advancing single image super-resolution.
• Networks with high effective parameter counts show enhanced robustness across diverse datasets and scale factors in SISR.
• The proposed REMA module offers a promising direction for improving detail preservation and overall quality in image super-resolution.