Lightweight Multi-Scale Asymmetric Attention Network for Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) have advanced single-image super-resolution (SISR) performance.
• Practical SISR is hindered by high computational costs and numerous parameters.

Purpose of the Study:

• To introduce a lightweight multi-scale asymmetric attention network (MAAN) for efficient SISR.
• To balance performance and parameter count in image super-resolution models.

Main Methods:

• The MAAN architecture comprises coarse-grained feature blocks (CFB), fine-grained feature blocks (FFB), and a reconstruction block (RB).
• Fine-grained feature blocks (FFB) utilize multi-scale attention residual blocks (MARB) to capture rich pixel-to-pixel correlations.
• Asymmetric multi-weights attention blocks (AMABs) within MARB generate attention maps to enhance SISR efficiency.

Main Results:

• The proposed MAAN demonstrates comparable performance to existing advanced lightweight SISR methods.
• MAAN achieves this performance with a significantly reduced number of parameters.
• Experimental results validate the effectiveness of the multi-scale and asymmetric attention mechanisms.

Conclusions:

• The lightweight MAAN offers an efficient solution for single-image super-resolution.
• The network design effectively balances SISR performance with computational efficiency.
• MAAN presents a promising approach for practical image super-resolution applications.